The Central Processing Unit has come a long way since the invention of the first digital, programmable computer in the middle of the 20th Century.
Here's a list of some of the more notable CPUs over the years.
open/close all folders
Pre-1970s
Vacuum tubes (1890s-1910s): The first electronic components.
Discrete circuts (1950s): Transistors, diodes (both previously vacuum tubes, now much smaller solid-state components), resistors, and capacitors wired together to build logic modules.
Logic chips (1960s): The first chips were simple one-operation logic modules, like digital LEGO bricks.
- Clock Speed: Tens to hundreds of KHz
- Word length: Varied enormously
- Address space: Up to 16 bits
- Used in: Early computers of the 1940s and '50s.
Discrete circuts (1950s): Transistors, diodes (both previously vacuum tubes, now much smaller solid-state components), resistors, and capacitors wired together to build logic modules.
- Clock Speed: Tens of KHz to 10 MHz
- Word length: Varied
- Address space: Usually 12-18 bits. Late in its life the PDP-11 used a 22-bit address space, allowing it to address 4 MB without page switching.
- Used in: 1950s and '60s Mainframes and Minicomputers, including the DEC PDP series.
- Magnavox Odyssey
Logic chips (1960s): The first chips were simple one-operation logic modules, like digital LEGO bricks.
- Clock Speed: Hundreds of KHz to tens of MHz (at the time they were popular; development has continued, and today they are in the GHz).
- Word length: Whatever you wanted. This is when word length began to be standardized using powers of 2 (8 bit/16 bit/32 bit etc.), but there were still a lot of other lengths in use, left over from earlier times. The DEC PDPs, for instance, used anywhere from 12 to 36 bit words.
- Address space: Again, whatever you wanted. The first computers with 32-bit address spaces appear here.
- Used in: 1960s and '70s mainframes and minicomputers.
- The first Atari arcade games, before 1976.
- Pong machines.
- Prototyping most of the CPUs, GPUs, sound chips, and other custom chips of the '70s and '80s.
- Widely used as 'glue logic' in consoles and computers from the 1970s onward; modern CMOS versions of the 7400 TTL chips are still produced and very common, as are the CD4000 chips (which were CMOS to begin with).
1970s
Intel 4004 (1971): First complete (CU and ALU) CPU.
Intel 8008 (1972): First 8-bit CPU.
Signetics 2650 (1973): An early, minicomputer-like CPU.
Intel 8080 (1974): First truly usable CPU, and the first to be used in hobbyist microcomputers, the ancestors of today's PCs.
Motorola 6800 (1974): First Motorola CPU.
MOS 6502 (1975): First cheap CPU. An improved Motorola 6800. The standard CPU for 8-bit videogame consoles and early Western home computers.
Zilog Z80 (1976): An improved Intel 8080. The main competitor to the MOS 6502, and more popular in Eastern countries.
Intel 8085 (1977): An improved 8080. While it, along with the 8080, was eclipsed by the Z80, it enjoyed success in industrial applications, including a radiation hardened version that NASA used in the 90s
Motorola 6809 (1978): Similar to the 6800, but with some 16-bit registers.
Intel 8086 (1978): The 16-bit evolution of the 8080. First in Intel's long line of x86 processors. The processor lacks a FPU, the FPU co-processor part, 8087, is sold as an option.
Motorola 68000 (1979): The first in Motorola's 68k family, and the first mainstream CPU to be capable of running 32-bit instructions. Internally, however, it was only a 16-bit design; Motorola officially termed it a 16/32-bit CPU, though most enthusiasts typically remember it as a 16-bit CPU, seeing how its best-remembered use cases were in home computers, consoles, and arcade machines that were promoted by their manufacturers as 16-bit (mostly because 32-bit instructions at the time weren't necessary or practical, given the heavy performance penalty, for those use cases).
- Clock Speed: 740 KHz
- Word length: 4 bits
- Address space: 12 bits
- Used in: Calculators; it was originally commissioned by a calculator manufacturer called Busicom.
Intel 8008 (1972): First 8-bit CPU.
- Clock Speed: 500-800 KHz
- Word length: 8 bits
- Address space: 14 bits
- Used in: Mainframe terminals; like the 4004, the 8008 was also a commission, this time by San Antonio terminal maker Datapoint. In an interesting twist, Datapoint had already designed the CPU and simply asked Intel to make a single-chip version of it; when this didn't pan out and Datapoint used their original discrete-component CPU, Intel got to keep the 8008, and the rest is history.
Signetics 2650 (1973): An early, minicomputer-like CPU.
- Clock Speed: 1.25-2(???) MHz
- Word length: 8 bits
- Address space: 15 bits
- Used in:
- Kit computers
- Central Data 2650
- Elektor TV Games Computer
- Emerson Arcadia 2001 (and clones)
- Interton VC 4000
- Arcade Games by Century Electronics, e.g. Hunchback
- Pinball tables by Zaccaria, e.g. Farfalla, Time Machine (Zaccaria)
Intel 8080 (1974): First truly usable CPU, and the first to be used in hobbyist microcomputers, the ancestors of today's PCs.
- Clock Speed: 2 MHz
- Word length: 8 bits
- Address space: 16 bits
- Used in:
- MITS Altair
- IMSAI 8080
- Some arcade games of the 1970s, including the Midway Games licensed version of Gun Fight (the earliest Arcade Game to use a microprocessor) and Space Invaders
Motorola 6800 (1974): First Motorola CPU.
- Clock Speed: 1-2 MHz
- Word length: 8 bits
- Address space: 16 bits
- Used in: Several hobbyist kit-computers, and the Sphere 1, the Trope Maker for the "Ctrl-Alt-Del" reset sequence on the IBM Personal Computer.
MOS 6502 (1975): First cheap CPU. An improved Motorola 6800. The standard CPU for 8-bit videogame consoles and early Western home computers.
- Clock Speed: 1-2 MHz
- Word length: 8 bits
- Address space: 16 bits
- Used in:
- Atari 2600: MOS 6507, a chopped-down version with a 13-bit address space.
- Atari 5200: Atari custom 6502C.
- Atari 7800: Atari custom 6502C.
- Atari Lynx: 4 MHz version.
- Nintendo Entertainment System: Ricoh RP2A03, a 6502 core with sound generators and some other extra features.
- TurboGrafx-16: Hudson Soft HuC6280, based on the Western Design Center 65C02, a faster version of the 6502 also equipped with a sound generator. The HuC6280 also saw use in a number of Data East arcade boards (including Bloody Wolf and Trio the Punch).
- Acorn System 1
- Apple ][ (and its kit-computer predecessor, the Apple I): Either the stock 6502 or the Western Design Center 65C02 (IIe Enhanced, IIc).
- Atari 8-Bit Computers: Either the stock 6502 or the Atari custom 6502C.
- BBC Micro (and all earlier Acorn microcomputers)
- Commodore PET
- Oric 1/Atmos
- VIC-20
- Commodore 64: MOS 6510, a 6502 with an extra register, used for bank switching.
- The Commodore 1541 Floppy Drive, one of the Commodore 64's most popular peripherals, used a 6502 as its controller and DOS processor.
- Commodore Plus/4: MOS 7501, a higher-clocked 6510.
- Commodore 128: MOS 8502, a faster and improved 6510, when running in Native Mode and C64 compatibility mode.
- SNES: The Sony SPC700, a 6502 derivative, was used as the processor in the sound module.
- The SPC700 has also shown up as a core in a bunch of unrelated Sony hardware, particularly VCRs.
- The SpaceShuttle
- Apple Macintosh IIfx: As an I/O coprocessor.
- Many arcade games by Atari, Data East, Exidy and Technos Japan. Irem and SNK used it for some of their earliest arcade games, but soon abandoned it in favor of the Z80.
Zilog Z80 (1976): An improved Intel 8080. The main competitor to the MOS 6502, and more popular in Eastern countries.
- Clock Speed: 2-6 MHz
- Word length: 8 bits
- Address space: 16 bits
- Used in:
- Astrocade
- Colecovision
- Game Boy: Sharp LR35902, a cross between an Intel 8080 and a Z80.
- Game Boy Color
- GX4000
- Sega SG-1000: NEC custom µPD780-1A.
- Sega Master System
- Game Gear
- Sega Genesis: As a coprocessor, and to provide backwards compatibility with the Master System.
- SNK Neo Geo: As a coprocessor
- Amstrad CPC
- Amstrad PCW
- MSX
- TRS-80
- ZX Spectrum
- Enterprise 64/128
- NEC PC-8801: NEC custom µPD780 or µPD70008.
- SAM Coupé
- Sharp MZ-700, MZ-800, MZ-1500, MZ-2000
- Sharp X1
- Commodore 128: As a 2nd CPU, for running in CP/M Mode. Unlike other systems, the Commodore 128 could not run both this and the MOS 8502 simultaneously (as in multiprocessing).
- Embedded systems and consumer electronics
- Arcade games from too many manufacturers to list; the Z80 has been used in thousands of arcade boards (counting those using it as a coprocessor)
- Electronic Music products such as Synthesizers, Drum Machines, and MIDI Sequencers.
Intel 8085 (1977): An improved 8080. While it, along with the 8080, was eclipsed by the Z80, it enjoyed success in industrial applications, including a radiation hardened version that NASA used in the 90s
- Clock Speed: 3, 5, and 6 MHz
- Word length: 8 bits
- Address space: 16 bits
- Used in:
- Tandy/Radio Shack TRS-80
- The Mars Pathfinder mission rover
- The Arcade Game Phoenix
Motorola 6809 (1978): Similar to the 6800, but with some 16-bit registers.
- Clock Speed: 1-2 MHz
- Word length: 8 bits
- Address space: 16 bits
- Used in:
- TRS80 & the Color Computer
- Dragon 32/64
- Fujitsu FM-7
- Vectrex
- Thomson MO/TO
- Many 1980s arcade games by Atari, Data East, Konami, Namco and Williams Electronics used either a 6809 or Hitachi's faster clone HD6309. Konami had their own custom encrypted versions as well.
Intel 8086 (1978): The 16-bit evolution of the 8080. First in Intel's long line of x86 processors. The processor lacks a FPU, the FPU co-processor part, 8087, is sold as an option.
- Clock Speed: 5-10 MHz
- Word length: 16 bits
- Address space: 20 bits
- Used in:
- IBM Personal Computer: Used the Intel 8088, comparable to the 8086 in most respects except for the external data bus being reduced from 16 bits to 8 bits. Some later "turbo" PCs such as the AT&T PC 6300 and IBM PS/2 Model 30 did have the actual 8086.
- IBM's DisplayWriter dedicated word processor
- NEC PC-9801 (earliest models)
- The Intel 8088 was the CPU in a few arcade games by Gottlieb, e.g. Q*bert.
- Noted for its quirk in how it calculated addresses. It added two numbers together to get the actual address without checking the results, so certain combinations resulted in a wrap around. Programmers took note of this, using the upper portion of memory for their program. When the address wrapped around, it pointed to the lowest portion, where most I/O memory was at. When its successor 80286 with its 24-bit address bus arrived, this wraparound trick didn't work anymore, breaking compatibility with the programs that used it. This bug was known as the A20 Line, after the 80286's 21-st address pin responsible for it.
Motorola 68000 (1979): The first in Motorola's 68k family, and the first mainstream CPU to be capable of running 32-bit instructions. Internally, however, it was only a 16-bit design; Motorola officially termed it a 16/32-bit CPU, though most enthusiasts typically remember it as a 16-bit CPU, seeing how its best-remembered use cases were in home computers, consoles, and arcade machines that were promoted by their manufacturers as 16-bit (mostly because 32-bit instructions at the time weren't necessary or practical, given the heavy performance penalty, for those use cases).
- Clock Speed: 8-20 MHz
- Word length: 32 bits internally, 16 bits externally
- Address space: 32 bits internally, 24 bits externally
- Used in:
- Atari Jaguar
- Philips CD-i: Custom version.
- Sega Genesis
- SNK Neo Geo
- Amiga 500, 600, 1000, 2000, Commodore CDTV
- Apple Macintosh
- Atari ST
- Sharp X68000
- Sinclair QL: Motorola 68008, with an 8-bit external data bus.
- TRS-80 Model 16
- Super A'Can
- Many arcade boards, e.g. Sega System 16, Capcom CP System I & II, later Toaplan games, earlier Cave games
- The ROSie Kiddie Ride control board also made use of the 68008 for running the control software.
1980s
Intel 80186 (1981): A slightly upgraded version of the 8086 (with an 8-bit version, the 80188), it integrated a lot of the external chips used with the 8086. It wasn't any faster per clock than its predecessor, meaning that most PC makers skipped over it and went straight to the 80286, but it saw wide use in embedded devices due to the integration of the other chips. Like the 8086 it does not have an FPU built in, there is an FPU co-processor, the 80187, that is sold separately.
Intel iAPX 432 (1981): Intel's first 32-bit processor, and the first of many attempts by the company to produce a viable VLIW architecture. Unfortunately, the design was way too ambitious for the manufacturing technologies available at the time, resulting in the CPU part alone having to be split into three chips, and forcing the system to be clocked relatively low in order to keep the three parts in sync. This, along with the architecture's poorly-designed compiler, resulted in performance around half that of the 8086, meaning that it was a failure pretty much out of the gate.
Intel 80286 (1982): The x86 processor that introduced Protected Mode, thanks to an integrated memory mapping unit (MMU) as well as Memory Addressing beyond 1MB. Its performance was double that of its predecessor. However, its design was somewhat rushed (due to Intel putting most of its resources toward the i432 "micro-mainframe"), and it had a number of quirks that had to be worked around for the sake of 8088 compatibility, particularly the "A20 line" issue (which was later put to good use on AT-class machines as the "high memory area"). Also, there was no way to return to real mode from protected mode, as the mindset was "Why would you want that?". The intended way if you really wanted to go back was a command from the keyboard that would reset the processor. However, Microsoft and others found it was much faster and cleaner to intentionally crash the CPU. This method is still used by some operating systems to intentionally reset the computer as a last resort. Like the 8086 and 80186 before it, 80286 lacks an FPU and has a FPU co-processor part, the 80287, that was sold separately.
Motorola 68020 (1984): A fully 32-bit 68000.
NEC V30 (1984): A pin-compatible clone of the Intel 8086. Despite the same clock speed could run code somewhat faster at due to improved internal logic. It also had a few extra features, including one that allowed it to emulate the Intel 8080.
Western Design Center 65816 (1984): A 16-bit derivative of the MOS 6502.
Berkley RISC (1984): The first RISC designed processor, developed as a government project from the 80s. It was commercialized soon afterwards as Sun's SPARC processor. When word spread out that RISC was very powerful (the second iteration outperformed Motorola's 68000 anywhere from 140% to 400%), many companies followed suit to build their own RISC chips. What came out of it was Intel's i960, AMD's 29000, DEC's Alpha, Motorola's 88000, and PowerPC.
Acorn Computers Acorn RISC Machine (ARM) (1985): One of the features that Acorn's BBC Micro included was something called "The Tube." It was essentially an external CPU socket, giving the computer some upgradeability and letting the original MOS 6502 it came with perform other tasks. After the success of the computer, the Micro's two designers looked to see what CPU they could add to it. The existing designs from Intel, National Semiconductor, and Motorola weren't satisfactory. However, designing their own CPU could be a herculean undertaking considering those companies were comparatively huge. But they had a look at one more chip, Western Design Center's 65C618, a 16-bit version of the 6502. The difference in this was that Western Design Center was a company with only a few engineers working out of an Arizona suburb. This gave Acorn the confidence that they could design their own chip.
So the team looked at what Berkeley did, and came up with their own design. They simplified a lot things from other desktop CPU designs, such as cutting the instruction count (45 total, compared to 357 for Intel's 80286), no on-die cache, no dedicated multiply/divide circuitry, and no floating point capabilities (though a barrel shifter helped). Also to help futureproof it, it was 32-bits from the get-go, which also meant that it didn't have to have circuitry to support 16-bit code and data, as well as being able to access memory directly since a 32-bit address space could cover any amount of RAM computers had at the time and well into the future. This meant systems didn't need to rely on things like segmenting memory (as in x86 or use bank switching like other 8-bit machines did.
Because of its simplicity, they were able to make an emulator and interpreter in BBC Micro BASIC, which they used to prove out the design before making the first chip. In some cases, they found ARM interpreted code to be faster than than natively compiled code. When Acorn got the first ARM chip and plugged it in, it worked on the first try. In addition, it consumed so little power (0.1W), it was able to work from the leakage current of test equipment. Performance-wise, it performed 10 times better than a similarly clocked Intel 80286 and as well as the 32-bit Motorola 68020 running at 17 MHz.
While the original chip didn't take off as much, it served to be such a promising first run that Acorn spun off this team into their own hardware company, eventually paving the way for future ARM processors.
* Clock Speed: 6 MHz
Intel 80386 (1985): The first 32-bit x86 processor, the 80386 also fixed several of the 80286's deficiencies; it could switch from protected mode to real mode without intentionally crashing the machine, and it supported 32-bit segments, meaning that the 80286's rather odd segmenting model could be avoided almost entirely. The 80386 also added "virtual 8086" mode, a way to run 16-bit code and 32-bit code simultaneously while in protected mode, which paved the way for Windows 3.x/9x's and 32-bit Windows NT's DOS box support. Came in two variants: the more powerful DX version (the original design), and the lower cost SX version with a 16-bit data bus (most motherboards at the time were 16-bit). Notably, the 387 had no FPU; this allowed manufacturers to sell the customers a base PC and offer an FPU (an 80387 or 80387SX) as an add-on.
Intel i960 (1985): Intel's second CPU with RISC architecture; mostly used as an embedded microcontroller and in military applications, and never promoted for general use.
MIPS R2000 (1985): First in the MIPS family of RISC CPUs. This was spawned off from a project from Stanford University to develop a RISC processor at the same time Berkeley was developing theirs.
Acorn ARM2 (1986): First in the ARM family of RISC CPUs, the most produced CPU family in history.
Motorola 68030 (1987): A 68020 with an integrated memory controller.
MIPS R3000 (1988)
Intel 80486/i486 (1989): First x86 processor with built-in level-1 cache and a built-in Floating Point Unit (except the SX variant, where it was a 486 with a faulty FPU), introduced Pipelining, and one of the first CPUs to have a multiplier. Due to its performance in games (notably with the FPU), it was in high demand in the early 90s. Also raised quite a stink when it was discovered that the 80487 FPU co-processor was essentially a fully-functional 80486DX with slight modifications made so it can't function without a 80486SX present, but can in fact work without it- when a 80487 is installed into a motherboard with an 80486SX CPU, the 80486SX gets disabled and the "80487" takes over completely. Initially came in speeds of up to 33MHz. Due to the controversy, the SX line was discontinued with the launch of the DX2 CPUs, offering speeds of up to 66MHz. Later, the DX4 CPUs was launched, capable of running at up to 100MHz.
Intel i860 (1989): Introduced along with the 80486, this was Intel's second attempt at a from-scratch RISC architecture. While in theory it was promising, it suffered from pipeline issues that impacted performance. Its floating point performance however was good to be used in several high-performance computing projects. Features implemented did influence later Intel designs as well. Allegedly, its codename (N-Ten) was where the name Windows NT came from.
NEC V60 (late 1980s): The first 32-bit general-purpose microprocessor mass-produced in Japan. Unlike NEC's earlier V20 and V30 processors, does not use the x86 architecture.
- Clock Speed: 5-25 MHz
- Word length: 16 bits
- Address space: 20 bits
- Used in: Tandy 2000, embedded systems
Intel iAPX 432 (1981): Intel's first 32-bit processor, and the first of many attempts by the company to produce a viable VLIW architecture. Unfortunately, the design was way too ambitious for the manufacturing technologies available at the time, resulting in the CPU part alone having to be split into three chips, and forcing the system to be clocked relatively low in order to keep the three parts in sync. This, along with the architecture's poorly-designed compiler, resulted in performance around half that of the 8086, meaning that it was a failure pretty much out of the gate.
- Clock Speed: 4-8 MHz
- Word length: 32 bits
- Address space: 40 bits
- Used in: Servers, mainframes
Intel 80286 (1982): The x86 processor that introduced Protected Mode, thanks to an integrated memory mapping unit (MMU) as well as Memory Addressing beyond 1MB. Its performance was double that of its predecessor. However, its design was somewhat rushed (due to Intel putting most of its resources toward the i432 "micro-mainframe"), and it had a number of quirks that had to be worked around for the sake of 8088 compatibility, particularly the "A20 line" issue (which was later put to good use on AT-class machines as the "high memory area"). Also, there was no way to return to real mode from protected mode, as the mindset was "Why would you want that?". The intended way if you really wanted to go back was a command from the keyboard that would reset the processor. However, Microsoft and others found it was much faster and cleaner to intentionally crash the CPU. This method is still used by some operating systems to intentionally reset the computer as a last resort. Like the 8086 and 80186 before it, 80286 lacks an FPU and has a FPU co-processor part, the 80287, that was sold separately.
- Clock Speed: 4-25 MHz
- Word length: 16 bits
- Address space: 24 bits
- Used in: IBM PCs and compatibles (starting with the IBM PC/AT).
Motorola 68020 (1984): A fully 32-bit 68000.
- Clock Speed: 16-50 MHz
- Word length: 32 bits
- Address space: 32 bits
- Used in:
- Amiga 1200, CD32: Both used the chopped-down 68EC020, with a 24-bit address space.
- Apple Macintosh II, LC
- Arcade Games: Atari GT (e.g. Primal Rage), Konami GX (e.g. Gokujou Parodius), Namco System 22 (e.g. Ridge Racer), Taito F3 (e.g. Bubble Symphony, Darius Gaiden), earlier Psikyo (e.g. Strikers 1945). Most of these used the 68EC020.
NEC V30 (1984): A pin-compatible clone of the Intel 8086. Despite the same clock speed could run code somewhat faster at due to improved internal logic. It also had a few extra features, including one that allowed it to emulate the Intel 8080.
- Clock Speed: 5-10 MHz
- Word length: 16 bits
- Address space: 20 bits
- Used in:
- Some "turbo" PCs came with, and the IBM Personal Computer could be upgraded with, the NEC V20, identical to the V30 except with the external data bus reduced to 8 bits to be pin-compatible with the 8088.
- NEC PC-9801
- NEC PC-88VA: NEC µPD9002, a V30-like processor compatible with the 8-bit CPU used in the PC-8801
- Wonderswan: Either V30 or the clone SPGY-1002.
- Arcade games by Irem (e.g. RType) and Seibu Kaihatsu (e.g. Raiden)
Western Design Center 65816 (1984): A 16-bit derivative of the MOS 6502.
- Clock Speed: 1-20 MHz
- Word length: 16 bits
- Address space: 24 bits
- Used in:
- Super Nintendo Entertainment System: Ricoh 5A22, a 65816 with some extra features.
- Apple IIGS
Berkley RISC (1984): The first RISC designed processor, developed as a government project from the 80s. It was commercialized soon afterwards as Sun's SPARC processor. When word spread out that RISC was very powerful (the second iteration outperformed Motorola's 68000 anywhere from 140% to 400%), many companies followed suit to build their own RISC chips. What came out of it was Intel's i960, AMD's 29000, DEC's Alpha, Motorola's 88000, and PowerPC.
- Processors in this family: Sun SPARC series of CPU
- Clock speed: 4+ MHz
- Word length: 32/64-bits (SPARC64/Berkley RISC V9 [=CPUs=] were 64-bit)
- Address space: 32/64-bits
- Used in: Sun Microsystems Sun-4 and SPARCStation, SPARCServer and SPARCCenter Workstations and Servers, several supercomputers.
- Otherwise was the blueprint for all following RISC architectures.
Acorn Computers Acorn RISC Machine (ARM) (1985): One of the features that Acorn's BBC Micro included was something called "The Tube." It was essentially an external CPU socket, giving the computer some upgradeability and letting the original MOS 6502 it came with perform other tasks. After the success of the computer, the Micro's two designers looked to see what CPU they could add to it. The existing designs from Intel, National Semiconductor, and Motorola weren't satisfactory. However, designing their own CPU could be a herculean undertaking considering those companies were comparatively huge. But they had a look at one more chip, Western Design Center's 65C618, a 16-bit version of the 6502. The difference in this was that Western Design Center was a company with only a few engineers working out of an Arizona suburb. This gave Acorn the confidence that they could design their own chip.
So the team looked at what Berkeley did, and came up with their own design. They simplified a lot things from other desktop CPU designs, such as cutting the instruction count (45 total, compared to 357 for Intel's 80286), no on-die cache, no dedicated multiply/divide circuitry, and no floating point capabilities (though a barrel shifter helped). Also to help futureproof it, it was 32-bits from the get-go, which also meant that it didn't have to have circuitry to support 16-bit code and data, as well as being able to access memory directly since a 32-bit address space could cover any amount of RAM computers had at the time and well into the future. This meant systems didn't need to rely on things like segmenting memory (as in x86 or use bank switching like other 8-bit machines did.
Because of its simplicity, they were able to make an emulator and interpreter in BBC Micro BASIC, which they used to prove out the design before making the first chip. In some cases, they found ARM interpreted code to be faster than than natively compiled code. When Acorn got the first ARM chip and plugged it in, it worked on the first try. In addition, it consumed so little power (0.1W), it was able to work from the leakage current of test equipment. Performance-wise, it performed 10 times better than a similarly clocked Intel 80286 and as well as the 32-bit Motorola 68020 running at 17 MHz.
While the original chip didn't take off as much, it served to be such a promising first run that Acorn spun off this team into their own hardware company, eventually paving the way for future ARM processors.
* Clock Speed: 6 MHz
- Word legth: 32-bits
- Address space: 32-bits capable, shiped with 26-bits
- Used in: Hobbyist kits and dev boards
Intel 80386 (1985): The first 32-bit x86 processor, the 80386 also fixed several of the 80286's deficiencies; it could switch from protected mode to real mode without intentionally crashing the machine, and it supported 32-bit segments, meaning that the 80286's rather odd segmenting model could be avoided almost entirely. The 80386 also added "virtual 8086" mode, a way to run 16-bit code and 32-bit code simultaneously while in protected mode, which paved the way for Windows 3.x/9x's and 32-bit Windows NT's DOS box support. Came in two variants: the more powerful DX version (the original design), and the lower cost SX version with a 16-bit data bus (most motherboards at the time were 16-bit). Notably, the 387 had no FPU; this allowed manufacturers to sell the customers a base PC and offer an FPU (an 80387 or 80387SX) as an add-on.
- Word length: 16 or 32 bits (DX version), 32 bits internally, 16 bits externally (SX version)
- Address space: 32 bits (DX version), 24 bits (SX version)
- Used in: IBM PCs and compatibles, NEC PC-9801, FM Towns (early models), FM Towns Marty (all models), the very first BlackBerry PDAs
Intel i960 (1985): Intel's second CPU with RISC architecture; mostly used as an embedded microcontroller and in military applications, and never promoted for general use.
- Clock Speed: 10-100 MHz
- Word length: 32 bits
- Address space: 33 bits
- Used in: Sega Model 2 arcade games (e.g. Daytona USA, Dead or Alive, Virtua Fighter 2), embedded systems
MIPS R2000 (1985): First in the MIPS family of RISC CPUs. This was spawned off from a project from Stanford University to develop a RISC processor at the same time Berkeley was developing theirs.
- Clock Speed: 8-15 MHz
- Word length: 32 bits
- Address space: 32 bits
- Used in: Silicon Graphics workstations
Acorn ARM2 (1986): First in the ARM family of RISC CPUs, the most produced CPU family in history.
- Clock Speed: 8-12 MHz
- Word length: 32 bits
- Address space: 26 bits
- Used in: Acorn Archimedes
Motorola 68030 (1987): A 68020 with an integrated memory controller.
- Clock Speed: 16-50 MHz
- Word Length: 32 bits
- Address Space: 32 bits
- Used in:
- Apple Macintosh SE/30, IIfx, Color Classic, LC II
- Amiga 3000 and 3000T
- Sharp X68030
MIPS R3000 (1988)
- Clock Speed: 20-33 MHz
- Word length: 32 bits
- Address space: 32 bits
- Used in: PlayStation
- PlayStation 2 for backwards compatibility with the PS1, as well as the I/O processor.
Intel 80486/i486 (1989): First x86 processor with built-in level-1 cache and a built-in Floating Point Unit (except the SX variant, where it was a 486 with a faulty FPU), introduced Pipelining, and one of the first CPUs to have a multiplier. Due to its performance in games (notably with the FPU), it was in high demand in the early 90s. Also raised quite a stink when it was discovered that the 80487 FPU co-processor was essentially a fully-functional 80486DX with slight modifications made so it can't function without a 80486SX present, but can in fact work without it- when a 80487 is installed into a motherboard with an 80486SX CPU, the 80486SX gets disabled and the "80487" takes over completely. Initially came in speeds of up to 33MHz. Due to the controversy, the SX line was discontinued with the launch of the DX2 CPUs, offering speeds of up to 66MHz. Later, the DX4 CPUs was launched, capable of running at up to 100MHz.
Contrary to popular belief, Intel could not trademark "486" because they simply failed to enforce it, not because you can't trademark numbers note . By the time they wanted to, other implementers of x86 processors were using the 486 moniker. So there wouldn't have been any legal ground for Intel to stand on to enforce any trademark. Thus for the next generation of processors, they figured they should start branding their processors with a trademarked name.
- Clock Speed: 16-100 MHz
- Word length: 16 or 32 bits
- Address space: 32 bits
- Used in: IBM PCs and compatibles, NEC PC-9801, FM Towns (later models)
Intel i860 (1989): Introduced along with the 80486, this was Intel's second attempt at a from-scratch RISC architecture. While in theory it was promising, it suffered from pipeline issues that impacted performance. Its floating point performance however was good to be used in several high-performance computing projects. Features implemented did influence later Intel designs as well. Allegedly, its codename (N-Ten) was where the name Windows NT came from.
- Clock speed: 20-50 MHz
- Word length: 32 bits (64-bit data path, 64-bit FPU)
- Address space: 32 bits
- Used in: Supercomputers, add-on coprocessors for high-performance PCs
NEC V60 (late 1980s): The first 32-bit general-purpose microprocessor mass-produced in Japan. Unlike NEC's earlier V20 and V30 processors, does not use the x86 architecture.
- Word length: 32 bits internally, 16 bits externally
- Address space: 24 bits
- Used in: Sega System 32, Sega Model 1 and SSV arcade games
1990s
Motorola 68040 (1990): Motorola's first 680x0 CPU with a built-in FPU. Faster than the i486 clock-per-clock, but ran notoriously hot (and thus was among the first desktop CPUs to require a heat sink). Came in several variants: the EC version (which dropped the FPU and MMU) used in Embedded Systems, and the low cost LC version (which dropped just the FPU; however this proved to be its undoing as buggy software searching for the FPU would crash the system).
MIPS R4000 (1991): First 64-bit RISC microprocessor.
Advanced RISC Machines ARM6 (1991)
Apple-IBM-Motorola PowerPC 601 & 603 (1992): The first two major forms of the PowerPC family. The 603 addressed and corrected most of the 601's flaws, and then fixed the remainder of them in its revised incarnation, the 603e. Still lacked official support for multiprocessing, though that didn't stop some resourceful designers.
DEC Alpha (1992): The first famous 64-bit CPU, and the champion of Clock Speed and floating-point performance throughout the 1990s. Was killed by Compaq after the DEC merger in favor of HP and Intel's then-upcoming Itanium CPU.
Intel Pentium (1993): The first superscalar x86 processor, with an integrated cache controller and a 64-bit data bus for faster memory access. Later models introduced the MMX instruction set. The name came from it being the 5th generation of the x86 processsor, in addition to likely making a marketable name, and more importantly, something that can be easily trademarked. Intel's marketing also grew quite prominent during this time such that within a few years, "Pentium" and Intel became a household name.
The processor however caused two major headaches for Intel. The first was the FDIV bug, which in rare circumstances, produced slightly incorrect results when performing floating point division. While this mostly affected academics and such who relied on accuracy, the headlines around this even in the 90s was enough that Intel was forced to offer a free replacement if requested. The second, which was discovered much later and on all versions of the initial Pentium architecture was the F00F bug. A specific instruction issued would cause the CPU to lock up. However, this one could easily be remedied in software and the Pentium II was out by the time this bug was discovered, so Intel wasn't on the hook to require replacements.
* Clock Speed: 60-300 MHz
Hitachi SH-2 (1993): Another low-cost RISC CPU. While its notable claim to fame is being used in video game systems (namely the Sega 32X and Saturn), it often found use in home entertainment electronics and engine control systems.
Advanced RISC Machines ARM7TDMI (1994): Gained fame as one of the most widely used embedded applications processors, and also the longest lasting. The instruction set was later licensed out (see ARM Cortex, below)
Motorola 68060 (1994): Motorola's last 680x0 CPU before venturing full time into the PowerPC family. Was actually more famously used in TV Character Generator systems. Like its predecessor the '040 (there was never a 68050), it came in EC and LC variants. Notably, Apple skipped this CPU entirely and went straight to the PowerPC.
NEC V810 (1994): Part of the V800 series of RISC CPUs.
MIPS R4300i (1995): An embedded systems variant of the MIPS R4200
Intel Pentium Pro (1995): RISC architecture with a CISC interpreter for the x86 ISA. Optimized for fully 32-bit OSes such as Windows NT and UNIX, where it was an excellent performer, but failed in the desktop market due to its high production cost and lackluster performance under Windows 3.x and 95 due to issues with it running 16-bit code. The high cost was mostly due to Intel looking into how to package processors going forward. CPUs were starting to require larger amounts of L2 cache to run optimally, but cache at the time already required its own chips and couldn't be integrated in the CPU itself. What Intel hoped to achieve with the Pentium Pro packaging was to bond all the L2 cache it needed on the same package. But due to Intel testing complete packages rather than the CPU and cache chips separately, if any one of them had problems, the entire thing had to be thrown away.
Cyrix 6x86 (1995): The first non-cloned x86 processor to pose a serious threat to Intel. Noted for its low price and excellent integer performance; however, its floating-point performance was lackluster, which became a problem as more games started to make use of FP code. It sold very well for its first two years, allowing Cyrix to take second place in the CPU market for a while. Unfortunately, Cyrix failed to significantly update the design (other than a relatively small refresh with the 6x86MX in 1998), meaning that it soon got left in the wake of the Celeron and K6-2 as the decade went on, and Cyrix ended up being purchased by VIA Technologies in the 2000s.
AMD K5 (1996): AMD's first x86 processor to be designed entirely in-house, having previously been a second-party manufacturer for Intel, and then having produced clones of Intel's designs. Like the Pentium Pro, it was a RISC design internally, with integer performance on-par with the 6x86 and floating-point performance almost as good as the Pentium. Unfortunately, AMD's manufacturing facilities couldn't keep up with their design teams, meaning that the K5 ended up late to market, and at Clock Speeds far too low to compete with the offerings from Intel or Cyrix.
Argonaut Technologies Limited (ATL) ARC (1996): The ARC processor is one of the many RISC designs that were created during the late 80s and throughout the 90s. Arguably a prototype of this processor started life on, of all things, Nintendo's Super Nintendo as the SuperFX chip. After building the chip, the hardware engineers at Argonaut Games decided to spin off into Argonaut Technologies Limited to develop the Argonaut RISC Core, or ARC. The company and the CPU design survived throughout the 2000s even, eventually being acquired by Snyopsys who continued to make updates all the way to 2020, releasing a third version of the ISA that brought it 64-bit support. It was likely retired in 2023 when Synopsis decided the ARC line would transition to the RISC-V architecture. But before then, the company claimed in 2014 that the processor was already in 1.4 billion devices.
Hitachi SH-4 (1997): An evolution to the SuperH family of processors. It was designed for high-performance use.
AIM PowerPC 740/750 (1997), aka PowerPC G3: An evolutionary derivative of the PowerPC 603e, the 740 was completely pin-compatible with the 603e and was therefore available from some after-market vendors as a drop-in upgrade. The major improvement that the 740 and 750 both had over the 603e was the addition of the PowerPC 604ev "Mach V" chip's extensive dynamic branch-prediction logic. However, because the G3 was based on the 603ev, it lacked the PowerPC 604's support for multiprocessing. Likewise, because it's based on the 603ev, it can be adopted to upgrade computers running the 603/604 CPUs.
AMD K6 (1997): AMD's first real challenge to Intel since the 486 days, and the beginning of its run as a Worthy Opponent. It started life as a design by a separate company named NexGen, who had a good design on paper but lacked the ability to produce it in any large quantities. AMD purchased NexGen and adapted the design to make it compatible with Pentium MMX motherboards while being almost as fast per-clock. The line eventually expanded into the K6-2 with "3DNow!" (floating-point SIMD) capability to make up for its somewhat weak standard FPU, and the rare but fast K6-III with built-in level-2 cache.
Cyrix MediaGX (1997): This featured the same CPU core as the 6x86, but added graphics, sound, memory and PCI controllers onto the very chip itself. While it was ahead of its time in many aspects, it had the bad luck of being launched when the 3D accelerator revolution was taking place, and the combination of a rather basic 2D graphics controller and the CPU's uninspiring 3D gaming performance meant it only found any usage in laptops and bargain-basement PCs. In retrospect, this chip may have been Cyrix's shining achievement; while the rest of their CPU line (and Cyrix itself) was purchased by Via for a relatively low price, this one actually got purchased by AMDnote , who continue to sell it to this day under the "Geode" name, albeit in the Geode GX and LX series (the NX series are rebranded Athlon XP processors).
Intel Pentium II (1997) and Pentium III (1999): Improved versions of the Pentium Pro, with the 16-bit performance problems fixed and a new economical cartridge design. In addition, they learned from the previous issue with the Pentium Pro in that they now tested the L2 cache and processor separately before attaching them to the cartridge itself. Despite this improvement, cache could only run at half the processor speed, so the goal of integrating cache into the CPU die itself remained on Intel's radar but couldn't do much about it until the next product lineup.
The Pentium III added "SSE" instructions for floating-point DSP work (following AMD's lead with "3DNow!"), and spawned a minor controversy over the use of embedded serial numbers that Intel eventually dropped. Later versions of the III were available in much smaller "flip-chip" packages, which were easier to install and cool. The Pentium II was made especially famous by what was probably Intel's craziest publicity gimmick in the form of their "Dancing Bunny Suit
" advertisements. This generation also saw the introduction of the Xeon line of server processors, which added multi-processor abilities,note usually also with more cache (and later, cores).
Intel Celeron (1998): A cheaper version of the Pentium II (and later III). The initial version (Covington) was heavily panned, due to it being essentially a Pentium II without its L2 cache. Meaning it couldn't outperform the Pentium w/ MMX at the same clock speed, though it found a niche with some overclockers who found it easy to tinker with.
The second version (Mendocino) however, pioneered the use of on-die L2 cache. While it only came with half the cache of its more expensive brethren, it ran at full bus speed whereas the Pentium II's L2 cache could only operate at half bus speed. This allowed it to perform almost as fast. becoming known as the Hypercompetent Sidekick of the CPU world. The line would continue until Intel's brand overhaul in 2023, serving as the cheapest version of the fully featured designs and sometimes as the "midrange" version of the ultra mobile design.
AMD Athlon & Athlon XP (1999): The processor that finally bested Intel, with much-improved floating-point performance over the K6 and a system bus design borrowed from the DEC Alpha to avoid legal issues with Intel. Later versions of the Athlon included the XP, MP (dual-processor capable) and value-priced Sempron. It was also the first processor to challenge the notion that clock speed meant performance. So much so that AMD named all of their processors with a "PR number", which supposedly represented the level it matched at against Pentium 4 at the speed of the number. The last version of the Athlon XP CPU was rebranded and sold under the Geode line as the Geode NX CPU to this day.
Sony Emotion Engine (1999): A custom processor based around the MIPS R5900. It implements MIPS III fully and MIPS IV partially. It included two co-processors that handed up to 4x32-bit vector calculations. However, it was regarded as tricky to use to its full potential, as the way to feed the co-processors required constantly feeding it data. In addition, one of them wouldn't be in use if you didn't use the other in the proper mode. Still, it had enough processing chops that the Japanese government banned its export without special permission because of fears it could be used in missile guidance systems.
Motorola PowerPC 7400/7410/744X/745X (1999): Motorola's solo venture in the PowerPC family. As IBM was falling back on designing the successor, Apple went with Motorola's design. It added multiprocessing support and a new feature that Motorola had branded AltiVec, the answer to Intel's SSE and AMD's 3DNow!. It was notably marketed by Apple, supposedly performing at 1GFLOP and being banned for export due to being a "weapons grade supercomputer". Indeed one demonstration that Apple referred to as "Debunking the Megahertz Myth," a side-by-side comparison was made of an operation on Adobe Photoshop running on an 867 MHz PowerMac G4 and on a 1.7 GHz Pentium 4 PC, with the G4 performing faster.
- Clock Speed: 20-40 MHz
- Word Length: 32 bit ALU, 64 bit FPU
- Address Space: 32 bits
- Used in:
- Apple Macintosh Quadra, Centris, Performa
- Amiga 4000 and 4000T
- NeXT
- Embedded Systems (EC variant)
MIPS R4000 (1991): First 64-bit RISC microprocessor.
- Clock Speed: 100 MHz
- Word length: 32 or 64 bits
- Address space: 36 bits
- Used in: PlayStation Portable
Advanced RISC Machines ARM6 (1991)
- Clock Speed: 12-33 MHz
- Word length: 32 bits
- Address space: 32 bits
- Used in:
- 3DO Interactive Multiplayer
- Acorn RiscPC
- Apple Newton
- Some Data East arcade boards (e.g. Captain America and the Avengers, Fighter's History)
Apple-IBM-Motorola PowerPC 601 & 603 (1992): The first two major forms of the PowerPC family. The 603 addressed and corrected most of the 601's flaws, and then fixed the remainder of them in its revised incarnation, the 603e. Still lacked official support for multiprocessing, though that didn't stop some resourceful designers.
- Clock Speed: 50-120 MHz (601), 75-100 MHz (603), 200-300 MHz (603e, 603ev)
- Word length: 32 bits
- Address space: 32 bits
- Used in:
- Power Macintosh (PCI)
- BeBox (multi-processing)
- Apple Pippin (underclocked 603)
- The PowerPC 602, a stripped-down version of the 603, was the CPU of 3DO's ill-fated M2 console.
DEC Alpha (1992): The first famous 64-bit CPU, and the champion of Clock Speed and floating-point performance throughout the 1990s. Was killed by Compaq after the DEC merger in favor of HP and Intel's then-upcoming Itanium CPU.
- Clock speed: 192 MHz-1.3 GHz
- Word length: 64 bits
- Address space: 64 bits
- Used in: DEC workstations and servers, supercomputers
Intel Pentium (1993): The first superscalar x86 processor, with an integrated cache controller and a 64-bit data bus for faster memory access. Later models introduced the MMX instruction set. The name came from it being the 5th generation of the x86 processsor, in addition to likely making a marketable name, and more importantly, something that can be easily trademarked. Intel's marketing also grew quite prominent during this time such that within a few years, "Pentium" and Intel became a household name.
The processor however caused two major headaches for Intel. The first was the FDIV bug, which in rare circumstances, produced slightly incorrect results when performing floating point division. While this mostly affected academics and such who relied on accuracy, the headlines around this even in the 90s was enough that Intel was forced to offer a free replacement if requested. The second, which was discovered much later and on all versions of the initial Pentium architecture was the F00F bug. A specific instruction issued would cause the CPU to lock up. However, this one could easily be remedied in software and the Pentium II was out by the time this bug was discovered, so Intel wasn't on the hook to require replacements.
* Clock Speed: 60-300 MHz
- Word length: 16 or 32 bits (64-bit memory access)
- Address space: 32 bits
- Used in: PCs, NEC PC-9821, FM Towns HB/HC
Hitachi SH-2 (1993): Another low-cost RISC CPU. While its notable claim to fame is being used in video game systems (namely the Sega 32X and Saturn), it often found use in home entertainment electronics and engine control systems.
- Clock Speed: ~29 MHz
- Word length: 32 bits
- Address space: 32 bits
- Used in:
- Sega 32X
- Sega Saturn
- Arcade games: Capcom CP System III (e.g. Red Earth, Street Fighter III), Sega Titan-V, later Psikyo games (e.g. Strikers 1945 II & III)
Advanced RISC Machines ARM7TDMI (1994): Gained fame as one of the most widely used embedded applications processors, and also the longest lasting. The instruction set was later licensed out (see ARM Cortex, below)
- Clock Speed: 17-70 MHz
- Word length: 32 bits
- Address space: 32 bits
- Used in:
- iPod classic
- Game Boy Advance
- Nintendo DS
Motorola 68060 (1994): Motorola's last 680x0 CPU before venturing full time into the PowerPC family. Was actually more famously used in TV Character Generator systems. Like its predecessor the '040 (there was never a 68050), it came in EC and LC variants. Notably, Apple skipped this CPU entirely and went straight to the PowerPC.
- Clock Speed: 50-75 MHz
- Word Length: 32 bits
- Address Space: 32 bits
- Used in:
- Character Generators
- Amiga 4000T
- PBX Systems
NEC V810 (1994): Part of the V800 series of RISC CPUs.
- Word Length: 32 bits
- Address Space: 32 bits
- Used in:
- NEC PC-FX
- Virtual Boy (customized version)
MIPS R4300i (1995): An embedded systems variant of the MIPS R4200
- Clock Speed: 100-133 MHz
- Word length: 32 or 64 bits
- Address space: 32 bits
- Used in:
- Nintendo 64: Licensed-built NEC VR4300
- SNK Hyper NeoGeo 64 hardware: Licensed-built NEC VR4300
Intel Pentium Pro (1995): RISC architecture with a CISC interpreter for the x86 ISA. Optimized for fully 32-bit OSes such as Windows NT and UNIX, where it was an excellent performer, but failed in the desktop market due to its high production cost and lackluster performance under Windows 3.x and 95 due to issues with it running 16-bit code. The high cost was mostly due to Intel looking into how to package processors going forward. CPUs were starting to require larger amounts of L2 cache to run optimally, but cache at the time already required its own chips and couldn't be integrated in the CPU itself. What Intel hoped to achieve with the Pentium Pro packaging was to bond all the L2 cache it needed on the same package. But due to Intel testing complete packages rather than the CPU and cache chips separately, if any one of them had problems, the entire thing had to be thrown away.
- Clock Speed: 150-200 MHz
- Word length: 16 or 32 bits (64-bit memory access)
- Address space: 36 bits
- Used in: PCs, particularly servers and high-end workstations
Cyrix 6x86 (1995): The first non-cloned x86 processor to pose a serious threat to Intel. Noted for its low price and excellent integer performance; however, its floating-point performance was lackluster, which became a problem as more games started to make use of FP code. It sold very well for its first two years, allowing Cyrix to take second place in the CPU market for a while. Unfortunately, Cyrix failed to significantly update the design (other than a relatively small refresh with the 6x86MX in 1998), meaning that it soon got left in the wake of the Celeron and K6-2 as the decade went on, and Cyrix ended up being purchased by VIA Technologies in the 2000s.
- Clock Speed: 75-250 MHz
- Word length: 16 or 32 bits (64-bit memory access)
- Address space: 32 bits
- Used in: PCs
AMD K5 (1996): AMD's first x86 processor to be designed entirely in-house, having previously been a second-party manufacturer for Intel, and then having produced clones of Intel's designs. Like the Pentium Pro, it was a RISC design internally, with integer performance on-par with the 6x86 and floating-point performance almost as good as the Pentium. Unfortunately, AMD's manufacturing facilities couldn't keep up with their design teams, meaning that the K5 ended up late to market, and at Clock Speeds far too low to compete with the offerings from Intel or Cyrix.
- Clock Speed: 75-133 MHz
- Word length: 16 or 32-bits (64-bit memory access)
- Address space: 32 bits
- Used in PCs
Argonaut Technologies Limited (ATL) ARC (1996): The ARC processor is one of the many RISC designs that were created during the late 80s and throughout the 90s. Arguably a prototype of this processor started life on, of all things, Nintendo's Super Nintendo as the SuperFX chip. After building the chip, the hardware engineers at Argonaut Games decided to spin off into Argonaut Technologies Limited to develop the Argonaut RISC Core, or ARC. The company and the CPU design survived throughout the 2000s even, eventually being acquired by Snyopsys who continued to make updates all the way to 2020, releasing a third version of the ISA that brought it 64-bit support. It was likely retired in 2023 when Synopsis decided the ARC line would transition to the RISC-V architecture. But before then, the company claimed in 2014 that the processor was already in 1.4 billion devices.
- Clock Speed: Varies
- Word length: 32-bits, later 64-bits
- Address space: 32 bits, later 64-bits
- Used in embedded systems
Hitachi SH-4 (1997): An evolution to the SuperH family of processors. It was designed for high-performance use.
- Clock Speed: 200 MHz
- Word length: 32 bits internally, 64 bits externally
- Address space: 32 bits
- Used in: Sega Dreamcast
AIM PowerPC 740/750 (1997), aka PowerPC G3: An evolutionary derivative of the PowerPC 603e, the 740 was completely pin-compatible with the 603e and was therefore available from some after-market vendors as a drop-in upgrade. The major improvement that the 740 and 750 both had over the 603e was the addition of the PowerPC 604ev "Mach V" chip's extensive dynamic branch-prediction logic. However, because the G3 was based on the 603ev, it lacked the PowerPC 604's support for multiprocessing. Likewise, because it's based on the 603ev, it can be adopted to upgrade computers running the 603/604 CPUs.
- Clock Speed: 200 MHz to 1 GHz
- Word length: 32 or 64 bits
- Address space: 32 bits
- Used in:
- Nintendo GameCube: Custom "Gekko" with extra instructions
- Wii: Custom "Broadway"
- PowerMacintosh G3, translucent iMac
- Third party kits for upgrading a 603/604 Macintosh to the G3 processor was made by several aftermarket manufacturers.
AMD K6 (1997): AMD's first real challenge to Intel since the 486 days, and the beginning of its run as a Worthy Opponent. It started life as a design by a separate company named NexGen, who had a good design on paper but lacked the ability to produce it in any large quantities. AMD purchased NexGen and adapted the design to make it compatible with Pentium MMX motherboards while being almost as fast per-clock. The line eventually expanded into the K6-2 with "3DNow!" (floating-point SIMD) capability to make up for its somewhat weak standard FPU, and the rare but fast K6-III with built-in level-2 cache.
- Clock speed: 166 MHz to 550 MHz
- Word size: 16 or 32 bits (64-bit memory access)
- Address space: 32 bits
- Used in: PCs
Cyrix MediaGX (1997): This featured the same CPU core as the 6x86, but added graphics, sound, memory and PCI controllers onto the very chip itself. While it was ahead of its time in many aspects, it had the bad luck of being launched when the 3D accelerator revolution was taking place, and the combination of a rather basic 2D graphics controller and the CPU's uninspiring 3D gaming performance meant it only found any usage in laptops and bargain-basement PCs. In retrospect, this chip may have been Cyrix's shining achievement; while the rest of their CPU line (and Cyrix itself) was purchased by Via for a relatively low price, this one actually got purchased by AMDnote , who continue to sell it to this day under the "Geode" name, albeit in the Geode GX and LX series (the NX series are rebranded Athlon XP processors).
- Clock Speed: 120 MHz to 300 MHz
- Word length: 16 or 32 bits (64-bit memory access)
- Address space: 32 bits
- Used in: PCs, laptops, embedded devices
Intel Pentium II (1997) and Pentium III (1999): Improved versions of the Pentium Pro, with the 16-bit performance problems fixed and a new economical cartridge design. In addition, they learned from the previous issue with the Pentium Pro in that they now tested the L2 cache and processor separately before attaching them to the cartridge itself. Despite this improvement, cache could only run at half the processor speed, so the goal of integrating cache into the CPU die itself remained on Intel's radar but couldn't do much about it until the next product lineup.
The Pentium III added "SSE" instructions for floating-point DSP work (following AMD's lead with "3DNow!"), and spawned a minor controversy over the use of embedded serial numbers that Intel eventually dropped. Later versions of the III were available in much smaller "flip-chip" packages, which were easier to install and cool. The Pentium II was made especially famous by what was probably Intel's craziest publicity gimmick in the form of their "Dancing Bunny Suit
" advertisements. This generation also saw the introduction of the Xeon line of server processors, which added multi-processor abilities,note usually also with more cache (and later, cores).- Clock Speed: 233 MHz to 1.4 GHz
- Word length: 16 or 32 bits (64-bit memory access)
- Address space: 36 bits
- Used in: Xbox, PCs
Intel Celeron (1998): A cheaper version of the Pentium II (and later III). The initial version (Covington) was heavily panned, due to it being essentially a Pentium II without its L2 cache. Meaning it couldn't outperform the Pentium w/ MMX at the same clock speed, though it found a niche with some overclockers who found it easy to tinker with.
The second version (Mendocino) however, pioneered the use of on-die L2 cache. While it only came with half the cache of its more expensive brethren, it ran at full bus speed whereas the Pentium II's L2 cache could only operate at half bus speed. This allowed it to perform almost as fast. becoming known as the Hypercompetent Sidekick of the CPU world. The line would continue until Intel's brand overhaul in 2023, serving as the cheapest version of the fully featured designs and sometimes as the "midrange" version of the ultra mobile design.
- Clock Speed: 266 MHz to 1.2 GHz
- Word length: 16 or 32 bits (64-bit memory access)
- Address space: 36 bits
- Used in: PCs
AMD Athlon & Athlon XP (1999): The processor that finally bested Intel, with much-improved floating-point performance over the K6 and a system bus design borrowed from the DEC Alpha to avoid legal issues with Intel. Later versions of the Athlon included the XP, MP (dual-processor capable) and value-priced Sempron. It was also the first processor to challenge the notion that clock speed meant performance. So much so that AMD named all of their processors with a "PR number", which supposedly represented the level it matched at against Pentium 4 at the speed of the number. The last version of the Athlon XP CPU was rebranded and sold under the Geode line as the Geode NX CPU to this day.
- Clock speed: 500 MHz to 1.4 GHz (Athlon)
- Clock speed: 1.33 GHz to 2.33 GHz (Athlon XP)
- Word size: 16 or 32 bits (64-bit memory access)
- Address space: 36 bits
- Used in: PCs, laptops. Also includes embedded devices after being repurposed as the Geode NX.
Sony Emotion Engine (1999): A custom processor based around the MIPS R5900. It implements MIPS III fully and MIPS IV partially. It included two co-processors that handed up to 4x32-bit vector calculations. However, it was regarded as tricky to use to its full potential, as the way to feed the co-processors required constantly feeding it data. In addition, one of them wouldn't be in use if you didn't use the other in the proper mode. Still, it had enough processing chops that the Japanese government banned its export without special permission because of fears it could be used in missile guidance systems.
- Clock Speed: ~300 MHz
- Word length: 128 bits, though external interfaces are either 16, 32, or 64 bits
- Address space: Probably 32 bits
- Used in:
- PlayStation 2
- Early PlayStation 3, for backward compatibility
Motorola PowerPC 7400/7410/744X/745X (1999): Motorola's solo venture in the PowerPC family. As IBM was falling back on designing the successor, Apple went with Motorola's design. It added multiprocessing support and a new feature that Motorola had branded AltiVec, the answer to Intel's SSE and AMD's 3DNow!. It was notably marketed by Apple, supposedly performing at 1GFLOP and being banned for export due to being a "weapons grade supercomputer". Indeed one demonstration that Apple referred to as "Debunking the Megahertz Myth," a side-by-side comparison was made of an operation on Adobe Photoshop running on an 867 MHz PowerMac G4 and on a 1.7 GHz Pentium 4 PC, with the G4 performing faster.
- Clock Speed: 350 MHz to 1.67 GHz, some aftermarket upgrades advertised as running at up to 2 GHz
- Word Length: 32 or 64 bits
- Address Space: Usually 32 bits
- Used In: Power Macintosh G4, PowerBook G4, iMac, iBook, early Mac Mini, Genesi Pegasos II, other PowerPC based Amiga clones.
- Unofficial upgrade kits for older AIM PowerPC G3 740/750 Macs are available from third party manufacturers like Sonnet Technologies.
2000s
Intel Pentium 4 Willamette and Northwood (2000-2002): Intel's first major blunder. The processor was marketed for its Clock Speed alone and the first generation, Willamette, performed on par, or worse than its predecessors. The high clock speed also made the processor run ridiculously hot as well. Intel's under-the-table deals kept AMD from winning overnight however, later causing Intel massive legal backlash. Things improved with the second generation, Northwood. Along with adding much faster bus and memory speeds, it introduced HyperThreading, which simulated a Multi-Core Processor. This combination led to it competing well against Athlon XP and even against the early Athlon 64s. Another small introduction in the Pentium 4 line was the Integrated Heat Spreader, a small metal cap which supposedly widened the heat dissipation area, though was more appreciated by enthusiasts for making it much harder to accidentally damage the processor while installing the heatsink. Later in this generation Intel also introduced the first enthusiast-oriented desktop CPU, the Pentium 4 Extreme Edition, which was a repackaged Xeon server CPU with more cache.
AMD Duron (2000): AMD's response to Intel's Celeron. Essentially a low cost Athlon.
Intel Itanium (2001): Intel's second major blunder. On paper it looked good with 64-bit technology, massive floating-point performance, the ability to access ludicrous amounts of memory, and back-compatibility with existing 32 bit code. In practice however, the first version barely even equaled a similarly-clocked Pentium III in native mode, and couldn't even convincingly outperform an 80486note in 32-bit mode. Subsequent editions were more successful, but it remained a very niche product due to its high price and the difficulty of generating optimal code for the architecture, and the Itanium gradually sank into irrelevancy as multi-core x86-64 chips equaled and surpassed its performance for a much lower price point. Intel finally pulled the plug on the family in 2021.
Transmeta Crusoe (2001): Perhaps more noted for the years of hype that surrounded its introduction, but noteworthy for being the first CPU designed from the ground up for laptops. Rather than using the CRISC design, the Crusoe used a Very Long Instruction Word (VLIW) design and had software to handle the instruction translation. Transmeta claimed that when their hardware and software became mature enough, they would be able to produce CPUs that were cheaper, much faster and much more power efficient than anything their rivals could make. Unfortunately, they never got the chance; after the Crusoe got a decidedly mixed reaction (it indeed had a very low power draw, but also pretty low performance, which is a problem when you're sitting around waiting for your laptop to do stuff on limited battery power), Intel, AMD and newcomers Via quickly moved into the "ultraportable laptop" market which the Crusoe had helped solidify, and Transmeta left the business in 2005.
Via C3 (2001): Via's debut in the CPU market, and a somewhat more successful implementation of what Transmeta had tried with the Crusoe. Initially this was branded the Cyrix III, but changed after troubles in the development processnote . Initially it was just another budget CPU design, but Via quickly emphasised its low power usage, creating very small CPU packages and the still widely-used Mini ITX case design to allow PCs to be built in the kind of space that had not previously been possible.
Loongson (2002): Like Russia with the MCST Elbrus 2000, China wanted a specific CPU that is designed in-country. The result is the MIPS-based Loongson. The first version of the CPU is based upon MIPS32, but were missing four instructions due to the CPU being unlicensed from MIPS Technology at that time. It was only finally licensed in 2007 for both MIPS32 and MIPS64 architectures. The Loongson II officially made the transition to 64-bit, and the Loongson III saw the introduction of multicore processing.
IBM PowerPC 970 (G5) (2003): The first 64-bit-capable PowerPC CPU intended for desktop use, the PowerPC 970 was announced by IBM in 2002. It was a cut-down, single-core version of IBM's POWER4 server CPU, with PowerPC compatibility and Motorola's AltiVec instructions added. Dubbed the "G5" by Apple, it first appeared in the Power Mac G5 in 2003, with single and dual-processor versions available. Apple was never quite pleased with the G5. While it was a powerful CPU, it was also a power hog and required extreme cooling measures because it ran so hot; some machines even shipped with water cooling stock, something that had never been needed before on a consumer-level machine. Apple lobbied IBM and Motorola to make a low-power version of the G5, something that would be suitable for the PowerBook, but neither was interested, forcing Apple to consider ARM and Intel. They went with Intel, and history was made; see the Core entry for more.
Intel Pentium M (2003): Intel's Ensemble Dark Horse during the age of the Pentium 4. This processor was designed from the ground up to be used in laptops, rather than using a scaled-down desktop part. Superficially it looks like a Pentium III with the Pentium 4's system bus and a huge cache bolted on, but every aspect of the chip was carefully hand-tuned to provide the best possible balance energy efficiency. As a result, it outperformed everything on the market in terms of performance-per-watt on release and humiliatingly for Intel it often outperformed the "Prescott" Pentium 4 models even at stock speeds. This led to companies producing desktop Pentium M boards for enthusiasts who wanted to get in on the act. The Pentium M would subsequently form the basis for the Core 2 line which eventually took the market back for Intel, and many believe that had Intel not had the Pentium M in development when they did, they would probably have never regained the performance lead.
AMD Athlon 64 and Athlon 64 X2 (2003): Introduced the x86-64 (also known as AMD64 and Intel EM64-T) instruction set, giving the x86 64-bit capabilities while remaining backwards compatible with 32-bit x86 programs. Also the first x86 processor to have the memory controller on-die, making access faster than it would be by going through the system bus. Noted for being by far the most successful AMD processor and probably remembered more fondly by enthusiasts than any other CPU, ever. The impact of its server counterpart, Opteron was even more keenly felt (64-bit addressing being more useful there), and AMD dominated that market for most of the decade. The dual-core Athlon 64 X2 was released in early 2005, and was the first fully integrated dual-core x86 CPU (the first overall, the Pentium D was released a few weeks earlier and was just two single-core Pentium 4s slapped together on the same package).
Intel Pentium 4 Prescott and Cedar Mill (2004): Intel's third major blunder, and widely regarded as their worst x86 processor. The design was an evolution of the previous generations, taking the pipeline and increasing its stages. However, this only made the heat and power problems of the Pentium 4 worse while also lowering per-clock performance, and caused Intel to cancel another generation (Tejas) which was intended to reach even higher clock speeds. This was the final blow to the Pentium brand, and Intel reduced it to being a budget chip (just above Celeron). The only significant changes elsewhere was that it introduced the Land Grid Array socket into Intel's lineup and added x86-64 in later versions. Most of the major issues were fixed with the core's second version, Cedar Mill, though that didn't last long on the market before Intel released its Core 2 line.
AMD Sempron (2004): The successor to the AMD Duron. Initially these were severely crippled Athlon 64s, with only one memory channel and no 64-bit support, but after the introduction of multi-core Athlon 64s the brand instead began to be used for AMD's single-core chips.
Intel Pentium D (2005): Technically the first dual-core x86 CPU, though it was just two processor dies in one package, whereas AMD's Athlon X2 was two cores in one die. There were two versions of the Pentium D; the first was "Smithfield," which was even more ridiculously hot than its Prescott forerunner and far slower than the Athlon 64 X2. The second version was "Presler," which didn't improve on a performance front, but had more manageable heat levels and actually won over some enthusiasts due to the insane overclocks that were feasible.
IBM Xenon (2005): Based on IBM's POWER4 architecture, this is actually a tri-core processor with each core being a modified PPE unit of the Cell Broadband Engine (see below).
Intel Core (2006): The point where Intel started to be taken seriously again. It evolved from the Pentium M, with other innovations included being able to downclock one of the two cores to save power, and a single cache that was shared by both cores. Notable also for Intel's debut on Apple Macintosh computers, which now enabled them to run much faster than the PowerPC series, and to the delight of many, be fully compatible with Microsoft Windows. A server version was also produced, but was generally ignored in this market due to the lack of 64-bit functionality; a problem which Intel would address quickly.
Sony-Toshiba-IBM Cell Broadband Engine (2006): A joint effort between the three companies, based on IBM's POWER4 architecture. Development started in 2001 as an ambitious effort to develop something to act like something between a general CPU and specialized, high performance processors (like GPUs). Basically it contains one general purpose CPU to control 8 smaller CPUs that do the real operations. However, since it emphasizes performance over anything else, it hasn't obtained wide success due to difficulty programming applications for it (although the proliferation of the PS3 would like one to think otherwise...)
Intel Core 2 (2006): Intel leap-frogs AMD after a 6-year dalliance with the inefficient and impractically hot-running Pentium 4. This processor went back to Intel's roots by overhauling the Pentium M and the original Core's design, making it faster, and adding AMD's x86-64 instructions. Only months later, Intel released the first quad-core desktop CPU, using the same dual-die approach previously employed for the Pentium D. In addition, Intel would continue to create lower power-consumption mobile variants for laptops. The Pentium brand, meanwhile, was reintroduced as a low-end CPU line, only a rung above the Celeron.
AMD Phenom, Phenom II and Athlon II (2007-2008): The successor to the Athlon 64, and the world's first native four-core chip. In addition to this, it was the first CPU that allowed the individual cores to be clocked wholly independently of one another as opposed to fixed rates, or even powered down altogether. AMD also offered tri-core versions in order to make up the production costs of chips which had a defective core. The first version of Phenom was late to the market, clocked far too slow to compete with Intel's offerings, and suffered glitches which could crash the whole system in certain rare circumstances.note However, the second revision was much more successful; while it couldn't convincingly outperform the Core 2 or later Core i7, AMD were able to offer more cores for the same price as an equivalent Intel chip, meaning that they dominated the low-end for several years. AMD also offered stripped-down versions of the Phenom and Phenom II, under the Athlon II brand; the Athlon II was mostly the same as its higher-end siblings, though with the third-level cache disablednote . Incidentally, this would be the last of AMD's CPU family to support their proprietary 3DNow! SIMD instructions. CPUs from AMD since lacks 3DNow! support save for two pre-cache instructions.
Intel Atom (2008): Years prior, Intel began shrinking it's mobile processors for use in Microsoft's push for tiny, handheld computers. However, Intel developed the Atom from scratch, rather than basing it off an existing part. When first revealed, the entire package itself was often compared with a penny and it sipped power (the chipset that it supported was larger and more power consuming). It started the Netbook phase of laptops and occasionally finds itself in "Nettops" (small computers, usually mounted on monitors). Currently Intel is using Atom to compete with ARM for the mobility and microserver sectors. After years of comparatively slow, incremental updates, Intel is poised to finally bring Atom up to speed with a schedule similar to the tick-tock cadence of the Core platform, as well as a general overhaul of the architecture, which includes bringing in integrated graphics (either PowerVR SGX or Intel HD Graphics), out-of-order execution and x86-64 instructions.
ARM Cortex (2008): Not so much a processor as it is an architectural design, as ARM does not make its own chips but licenses them instead. It is based on the ARM7TDMI instruction set mentioned earlier. However, it found itself being one of the most widely used processors in low-power electronics such as embedded devices, Netbook and Nettop computers, smartphones, and portable game systems. Chances are, if it's not using x86, it's probably using an ARM Cortex design. The family comes in three flavors: (A)pplication, (M)icrocontroller, and (R)obust. It is later upgraded to 64-bit (see ARMv8-A, below)
Intel Nehalem & Westmere (2008-2010): Intel started to implement many features of then-current and previous x86 processors, including on-die memory controllers, HyperThreading, and single die quad-core. Models with a K at the end had their clock multipliers unlocked that was within many enthusiasts' budgets, rather than being a feature on the "Extreme Edition" parts that commanded luxury prices. New to Intel's lineup was an overclock-as-needed feature called Turbo Boost and graphics on the processor package itself for dual-core models. One of two issues was that Intel had an unclear branding schema, particularly with the i5 series note . The other was that motherboard sockets were split between markets, which drew criticism among AMD fans note , and had a staggered release, with the mainstream following the high end by a year. note Also of note: the very first consumer 6-core models note .
First Generation MCST Elbrus 2000 (1999-2011)A Russian attempt to revive its struggling microelectronics industry, this CPU was in development since early 2000, but was largely sidelined by a lack of modern production facilities, as Russian laws at the time prevented its large-scale production on the foreign plants, and their major application was in military. Another attempt to create a competitive VLIW-based CPU, this looks like it was finally done right: back in 2005 one of this chips, clocked at 300 MHz, outpaced a 500 MHz Pentium III in x86 compatibility mode, and was competitive with 2 GHz P4 when running a native code. As common for the VLIW CPU internally it is split into several pipelined execution units that are directly driven by a command stream without costly transcoding and rearranging the complex CISC and even RISC commands. Its secret, however, lies in the extremely efficient compile algorithms that allow for correct arrangement of this stream. It also emulates x86 architecture by a special binary compiler that runs as a hardware-supported shadow task in an OS and translates x86 binaries on-the-fly with a minimal penalty.
- Clock Speed: 1.2 GHz to 3.46GHz
- Word length: 16 or 32 bits
- Address space: 36 bits
- Used in: PCs
AMD Duron (2000): AMD's response to Intel's Celeron. Essentially a low cost Athlon.
- Clock Speed: 600 MHz to 1.8 GHz
- Word size: 16 or 32 bits (64-bit memory access)
- Address space: 36 bits
- Used in: PCs
Intel Itanium (2001): Intel's second major blunder. On paper it looked good with 64-bit technology, massive floating-point performance, the ability to access ludicrous amounts of memory, and back-compatibility with existing 32 bit code. In practice however, the first version barely even equaled a similarly-clocked Pentium III in native mode, and couldn't even convincingly outperform an 80486note in 32-bit mode. Subsequent editions were more successful, but it remained a very niche product due to its high price and the difficulty of generating optimal code for the architecture, and the Itanium gradually sank into irrelevancy as multi-core x86-64 chips equaled and surpassed its performance for a much lower price point. Intel finally pulled the plug on the family in 2021.
- Clock Speed: 733 MHz to 1.6GHz
- Word length: 64 bits (32 bits via hardware emulation; later editions swapped this for a software emulator)
- Address space: 50 bits
- Used in: Servers, supercomputers
Transmeta Crusoe (2001): Perhaps more noted for the years of hype that surrounded its introduction, but noteworthy for being the first CPU designed from the ground up for laptops. Rather than using the CRISC design, the Crusoe used a Very Long Instruction Word (VLIW) design and had software to handle the instruction translation. Transmeta claimed that when their hardware and software became mature enough, they would be able to produce CPUs that were cheaper, much faster and much more power efficient than anything their rivals could make. Unfortunately, they never got the chance; after the Crusoe got a decidedly mixed reaction (it indeed had a very low power draw, but also pretty low performance, which is a problem when you're sitting around waiting for your laptop to do stuff on limited battery power), Intel, AMD and newcomers Via quickly moved into the "ultraportable laptop" market which the Crusoe had helped solidify, and Transmeta left the business in 2005.
- Clock Speed: 500 MHz to 800 MHz
- Word length: 64 bits (16 and 32 bits via software emulation)
- Address space: 32 bits
- Used in: Laptops
Via C3 (2001): Via's debut in the CPU market, and a somewhat more successful implementation of what Transmeta had tried with the Crusoe. Initially this was branded the Cyrix III, but changed after troubles in the development processnote . Initially it was just another budget CPU design, but Via quickly emphasised its low power usage, creating very small CPU packages and the still widely-used Mini ITX case design to allow PCs to be built in the kind of space that had not previously been possible.
- Clock Speed: 500 MHz to 1.2 GHz
- Word length: 16 or 32 bits (64 internally)
- Address space: 32 bits
- Used in: Laptops, Small Form Factor PCs
Loongson (2002): Like Russia with the MCST Elbrus 2000, China wanted a specific CPU that is designed in-country. The result is the MIPS-based Loongson. The first version of the CPU is based upon MIPS32, but were missing four instructions due to the CPU being unlicensed from MIPS Technology at that time. It was only finally licensed in 2007 for both MIPS32 and MIPS64 architectures. The Loongson II officially made the transition to 64-bit, and the Loongson III saw the introduction of multicore processing.
- Clock speed: 200 MHz to 1.5 GHz
- Core Count: 1 (Loongson/Loongson II) to 64 (Godsun-T. This is not a typo.)
- Word length: 32 bits (first generation Loongson and Godson-T), 64 bits (Loongson II/III)
- Address space: Unknown
- Fab process: 180nm (first generation Loongson) to 28nm (Loongson III)
- Used in: Laptops and Mini-PCs from a Chinese company called Lemote, EMTEC gdium notebook, Dawning 6000 supercomputer.
IBM PowerPC 970 (G5) (2003): The first 64-bit-capable PowerPC CPU intended for desktop use, the PowerPC 970 was announced by IBM in 2002. It was a cut-down, single-core version of IBM's POWER4 server CPU, with PowerPC compatibility and Motorola's AltiVec instructions added. Dubbed the "G5" by Apple, it first appeared in the Power Mac G5 in 2003, with single and dual-processor versions available. Apple was never quite pleased with the G5. While it was a powerful CPU, it was also a power hog and required extreme cooling measures because it ran so hot; some machines even shipped with water cooling stock, something that had never been needed before on a consumer-level machine. Apple lobbied IBM and Motorola to make a low-power version of the G5, something that would be suitable for the PowerBook, but neither was interested, forcing Apple to consider ARM and Intel. They went with Intel, and history was made; see the Core entry for more.
- Clock speed: 1.6-2.7 GHz
- Word length: 64 bits
- Address space: 64 bits
- Used in: Power Mac G5, iMac G5, IBM servers
Intel Pentium M (2003): Intel's Ensemble Dark Horse during the age of the Pentium 4. This processor was designed from the ground up to be used in laptops, rather than using a scaled-down desktop part. Superficially it looks like a Pentium III with the Pentium 4's system bus and a huge cache bolted on, but every aspect of the chip was carefully hand-tuned to provide the best possible balance energy efficiency. As a result, it outperformed everything on the market in terms of performance-per-watt on release and humiliatingly for Intel it often outperformed the "Prescott" Pentium 4 models even at stock speeds. This led to companies producing desktop Pentium M boards for enthusiasts who wanted to get in on the act. The Pentium M would subsequently form the basis for the Core 2 line which eventually took the market back for Intel, and many believe that had Intel not had the Pentium M in development when they did, they would probably have never regained the performance lead.
- Clock Speed: 1 GHz to 2.26 GHz
- Word length: 16 or 32 bits (64 internally)
- Address space: 32 bits
- Used in: Laptops, some PCs
AMD Athlon 64 and Athlon 64 X2 (2003): Introduced the x86-64 (also known as AMD64 and Intel EM64-T) instruction set, giving the x86 64-bit capabilities while remaining backwards compatible with 32-bit x86 programs. Also the first x86 processor to have the memory controller on-die, making access faster than it would be by going through the system bus. Noted for being by far the most successful AMD processor and probably remembered more fondly by enthusiasts than any other CPU, ever. The impact of its server counterpart, Opteron was even more keenly felt (64-bit addressing being more useful there), and AMD dominated that market for most of the decade. The dual-core Athlon 64 X2 was released in early 2005, and was the first fully integrated dual-core x86 CPU (the first overall, the Pentium D was released a few weeks earlier and was just two single-core Pentium 4s slapped together on the same package).
- Clock Speed: 1.8 GHz to 3.5 GHz
- Core count: 1/2
- Word length: 16, 32 or 64 bits; built-in memory controller
- Address space: 36 or 64 bits (limited to 48 due to practicality)
- Used in: PCs
Intel Pentium 4 Prescott and Cedar Mill (2004): Intel's third major blunder, and widely regarded as their worst x86 processor. The design was an evolution of the previous generations, taking the pipeline and increasing its stages. However, this only made the heat and power problems of the Pentium 4 worse while also lowering per-clock performance, and caused Intel to cancel another generation (Tejas) which was intended to reach even higher clock speeds. This was the final blow to the Pentium brand, and Intel reduced it to being a budget chip (just above Celeron). The only significant changes elsewhere was that it introduced the Land Grid Array socket into Intel's lineup and added x86-64 in later versions. Most of the major issues were fixed with the core's second version, Cedar Mill, though that didn't last long on the market before Intel released its Core 2 line.
- Clock Speed: 2.4 GHz to 3.8GHz
- Word length: 16, 32 or 64 bits
- Address space: 42 bits
- Used in: PCs
AMD Sempron (2004): The successor to the AMD Duron. Initially these were severely crippled Athlon 64s, with only one memory channel and no 64-bit support, but after the introduction of multi-core Athlon 64s the brand instead began to be used for AMD's single-core chips.
- Clock Speed: 1-2.9 GHz
- Word Length: 16, 32 or 64 bits (32-bit only on earlier versions)
- Address Space: 36 or 64 bits (limited to 48 due to practicality)
- Used in: PCs
Intel Pentium D (2005): Technically the first dual-core x86 CPU, though it was just two processor dies in one package, whereas AMD's Athlon X2 was two cores in one die. There were two versions of the Pentium D; the first was "Smithfield," which was even more ridiculously hot than its Prescott forerunner and far slower than the Athlon 64 X2. The second version was "Presler," which didn't improve on a performance front, but had more manageable heat levels and actually won over some enthusiasts due to the insane overclocks that were feasible.
- Clock Speed: 2.66 GHz to 3.73GHz (some manufacturers offered a semi-official 4.26GHz version of "Presler")
- Word length: 16, 32 or 64 bits
- Address space: 42 bits
- Used in: PCs
IBM Xenon (2005): Based on IBM's POWER4 architecture, this is actually a tri-core processor with each core being a modified PPE unit of the Cell Broadband Engine (see below).
- Clock Speed: 3.2 GHz
- Word length: 32 or 64 bits
- Address space: Probably 32 or 42 bits
- Used in: Xbox 360
Intel Core (2006): The point where Intel started to be taken seriously again. It evolved from the Pentium M, with other innovations included being able to downclock one of the two cores to save power, and a single cache that was shared by both cores. Notable also for Intel's debut on Apple Macintosh computers, which now enabled them to run much faster than the PowerPC series, and to the delight of many, be fully compatible with Microsoft Windows. A server version was also produced, but was generally ignored in this market due to the lack of 64-bit functionality; a problem which Intel would address quickly.
- Clock Speed: 1.0 to 2.13 GHz
- Word length: 16 or 32 bits
- Address space: 36 bits
- Used in: Laptops, early Intel iMacs
Sony-Toshiba-IBM Cell Broadband Engine (2006): A joint effort between the three companies, based on IBM's POWER4 architecture. Development started in 2001 as an ambitious effort to develop something to act like something between a general CPU and specialized, high performance processors (like GPUs). Basically it contains one general purpose CPU to control 8 smaller CPUs that do the real operations. However, since it emphasizes performance over anything else, it hasn't obtained wide success due to difficulty programming applications for it (although the proliferation of the PS3 would like one to think otherwise...)
- Clock Speed: 3.2 GHz
- Core count: 1 PPE, 8 SPUs
- Word length: 32 or 64 bits
- Address space: 64 bits
- Used in: PlayStation 3, IBM BladeCenter, IBM Roadrunner supercomputer note
- There is a case where the Cell was used in a media encoding add-in card that performed really well for the task.
Intel Core 2 (2006): Intel leap-frogs AMD after a 6-year dalliance with the inefficient and impractically hot-running Pentium 4. This processor went back to Intel's roots by overhauling the Pentium M and the original Core's design, making it faster, and adding AMD's x86-64 instructions. Only months later, Intel released the first quad-core desktop CPU, using the same dual-die approach previously employed for the Pentium D. In addition, Intel would continue to create lower power-consumption mobile variants for laptops. The Pentium brand, meanwhile, was reintroduced as a low-end CPU line, only a rung above the Celeron.
- Clock speed: 1.0 to 3.4 GHz
- Core count: 2/4 (Core 2 Quad)
- Word length: 16, 32 or 64 bits
- Address space: 36 or 64 bits (with the same 48-bit caveat as the Athlon 64)
- Used in: PCs, lower end models in Sega's arcade systems.
AMD Phenom, Phenom II and Athlon II (2007-2008): The successor to the Athlon 64, and the world's first native four-core chip. In addition to this, it was the first CPU that allowed the individual cores to be clocked wholly independently of one another as opposed to fixed rates, or even powered down altogether. AMD also offered tri-core versions in order to make up the production costs of chips which had a defective core. The first version of Phenom was late to the market, clocked far too slow to compete with Intel's offerings, and suffered glitches which could crash the whole system in certain rare circumstances.note However, the second revision was much more successful; while it couldn't convincingly outperform the Core 2 or later Core i7, AMD were able to offer more cores for the same price as an equivalent Intel chip, meaning that they dominated the low-end for several years. AMD also offered stripped-down versions of the Phenom and Phenom II, under the Athlon II brand; the Athlon II was mostly the same as its higher-end siblings, though with the third-level cache disablednote . Incidentally, this would be the last of AMD's CPU family to support their proprietary 3DNow! SIMD instructions. CPUs from AMD since lacks 3DNow! support save for two pre-cache instructions.
- Clock Speed: 1.8 GHznote to 3.3 GHznote (up to 3.7 GHz with Turbo).
- Core count: 2/3/4, Phenom II added 6-core variants (Phenom II 10xx series).
- Word length: 16, 32 or 64 bits; built-in memory controller
- Address space: 36 or 64 bits (limited to 48 due to practicality)
- Used in: PCs
Intel Atom (2008): Years prior, Intel began shrinking it's mobile processors for use in Microsoft's push for tiny, handheld computers. However, Intel developed the Atom from scratch, rather than basing it off an existing part. When first revealed, the entire package itself was often compared with a penny and it sipped power (the chipset that it supported was larger and more power consuming). It started the Netbook phase of laptops and occasionally finds itself in "Nettops" (small computers, usually mounted on monitors). Currently Intel is using Atom to compete with ARM for the mobility and microserver sectors. After years of comparatively slow, incremental updates, Intel is poised to finally bring Atom up to speed with a schedule similar to the tick-tock cadence of the Core platform, as well as a general overhaul of the architecture, which includes bringing in integrated graphics (either PowerVR SGX or Intel HD Graphics), out-of-order execution and x86-64 instructions.
- Clock Speed: 600MHz to 2.13GHz
- Word length: 16, 32 or 64-bit
- Address space: 32-bit/48-bit
- Used in: Netbooks, Nettops, tablets, and microservers.
ARM Cortex (2008): Not so much a processor as it is an architectural design, as ARM does not make its own chips but licenses them instead. It is based on the ARM7TDMI instruction set mentioned earlier. However, it found itself being one of the most widely used processors in low-power electronics such as embedded devices, Netbook and Nettop computers, smartphones, and portable game systems. Chances are, if it's not using x86, it's probably using an ARM Cortex design. The family comes in three flavors: (A)pplication, (M)icrocontroller, and (R)obust. It is later upgraded to 64-bit (see ARMv8-A, below)
- Notable implementations: Apple's A4 through A6X, Texas Instrument's OMAP, many Qualcomm's Snapdragons, NVIDIA's Tegra except the K1 Denver, and many Samsung's Exynos. Also found in many bargain-basement devices that uses CPUs from RockChip, AllWinner, MediaTek and HiSilicon.
- Clock Speed: None defined (implementations as of 2013 reached 2.5GHz)
- Word length: 32-bit
- Address space: 32-bit
- Used in (notably):
- Apple's iPhone (up until iPhone 5c, the iPhone 5s uses ARMv8-A instead), iPod Touch, and iPad (up until the original iPad Mini)
- Most Android based devices
- Sony's Play Station Vita
- Nintendo DS and Nintendo 3DS.
Intel Nehalem & Westmere (2008-2010): Intel started to implement many features of then-current and previous x86 processors, including on-die memory controllers, HyperThreading, and single die quad-core. Models with a K at the end had their clock multipliers unlocked that was within many enthusiasts' budgets, rather than being a feature on the "Extreme Edition" parts that commanded luxury prices. New to Intel's lineup was an overclock-as-needed feature called Turbo Boost and graphics on the processor package itself for dual-core models. One of two issues was that Intel had an unclear branding schema, particularly with the i5 series note . The other was that motherboard sockets were split between markets, which drew criticism among AMD fans note , and had a staggered release, with the mainstream following the high end by a year. note Also of note: the very first consumer 6-core models note .
- Processors in this family: Celeron, Pentium, Core i3, Core i5, Core i7
- Clock speed: 2.66 to 3.6 GHz
- Core count: 2/4/6
- Word length: 16, 32 or 64 bits
- Address space: 36 or 64 bits (48-bit in practice)
- Fab process: 45nm (Nehalem), 32nm (Westmere)
- Used in: PCs
First Generation MCST Elbrus 2000 (1999-2011)A Russian attempt to revive its struggling microelectronics industry, this CPU was in development since early 2000, but was largely sidelined by a lack of modern production facilities, as Russian laws at the time prevented its large-scale production on the foreign plants, and their major application was in military. Another attempt to create a competitive VLIW-based CPU, this looks like it was finally done right: back in 2005 one of this chips, clocked at 300 MHz, outpaced a 500 MHz Pentium III in x86 compatibility mode, and was competitive with 2 GHz P4 when running a native code. As common for the VLIW CPU internally it is split into several pipelined execution units that are directly driven by a command stream without costly transcoding and rearranging the complex CISC and even RISC commands. Its secret, however, lies in the extremely efficient compile algorithms that allow for correct arrangement of this stream. It also emulates x86 architecture by a special binary compiler that runs as a hardware-supported shadow task in an OS and translates x86 binaries on-the-fly with a minimal penalty.
- Clock Speed: 300-500 MHz
- Core Count: 1 - 4
- Word Length: 32 or 64 bits
- Address Space: 64 bits (see above)
- Fab Process: 250 and 130 nm
- Instruction Set: Native or x86 compatibility mode through a binary translation
- Used In: Military computers.
2010s
Intel Sandy Bridge & Ivy Bridge (2011): Intel starts to focus on power efficiency and the use of integrated graphics to accelerate compute tasks. Ivy Bridge was the first processor to use Intel's new tri-gate transistors which increased power efficiency and allowed them to create the Ultrabook initiative, PCs inspired after Apple's Macbook Air. Carrying over from the previous generation, Intel also released K versions of the processor with unlocked clock multipliers. The high-end sockets, simply denoted by an -E suffix, left off the graphics but packed up to six cores. Ivy Bridge also used thermal paste in the heatspreader, which causes higher temperatures, but at the same time allows the heatspreader to be removed more easily. Previously they were soldered on.
AMD Fusion (Bobcat, Llano, Jaguar) (2011): After AMD bought ATi in 2006, AMD had goals to merge the logical portion of the processor, with the computational powerhouse of the GPU. AMD eventually came out with a similar solution to that of Intel's in 2011 named Fusion. AMD's solution has more of a GPU focus, die shots showing that AMD's "Accelerated Processing Units" (APU) are roughly half GPU and half CPU. These features are rather popular with laptops, as graphical and computing performance increases dramatically, with little impact to battery life. In late 2012, the updated Trinity series of APUs was released, finally making the line viable for desktop users.
AMD FX-series (2011): AMD takes a radical approach to CPU organization in the processor. Instead of a "core", this processor has "modules", consisting of two basic execution units, with a shared FPU. There was claim that this used die space more efficiently. While the first version, "Bulldozer" did excel in multi-threaded tasks (up to the same performance as Intel's second generation Core i7 in some cases), in other tasks, it's only slightly better (or even slightly slower) than its predecessor. The second version, "Piledriver" improved matters — especially after AMD released the 9000-series with ludicrously high clock speeds — consistently equalling or beating the equivalent Core i7s in heavily threaded situations, albeit at the cost of much higher power consumption (about 2.5x that of Intel's offerings) and still-low single thread performance. However, AMD conceded that the line had no long-term future, and pulled the plug on FX versions of the next two planned cores, Steamroller and Excavator (though both found their way into their APUs). As of 2014, the series was repurposed as a part of the branding used on Fusion APUs, with the upcoming two FX chips being APUs instead of being discreet CPU-only chips.
Second Generation Elbrus 2000 (2011)An updated e2k CPU with smaller tech norms and increased clock speed. Comparable to low-end Intel chips clocked several times faster in x86 emulation mode, and lightning-fast on the native VLIW code. Also extremely cool for its performance, radiating ~15-20W at full load. Recent versions also included several DSP cores to speed-up the image and signal processing, a vestige of its previous use as a radar CPU. While still mainly used in military applications, MCST has recently demonstrated a PC-standard mini-ITX motheboard, indicating their wish to enter the general PC market, especially multimedia servers, where their additional DSP cores would be especially useful.
Advanced RISC Machines ARMv8-A (2011): ARM goes 64-bit. With the demand for multitasking smartphones and tablets on the rise, and the demand for CPU-intensive gaming going mainstream, it is starting to become clear that the 17-year-old 32-bit ARM Cortex(ARM7TDMI-derived) instruction set will not be sufficient. This evolution promised that the ARM CPU will be able to handle more precise (and larger) numbers and address more memory, something important with large games.
Intel Xeon Phi (2012): The Xeon Phi started its roots in the ambitious GPU Larabee back in 2007, which Intel wanted it to do real-time ray tracing (the Holy Grail of real-time graphics, which in the end would not be achieved until nVidia released their Turing GPU in 2018). However, this fell through and was brought up on and off in 2010. The end result though, is what Intel dubbed the Many-Integrated Core (MIC) processor for highly parallel processing. Using a revamped P5 architecture (the same one in the original Pentium), the low-end Phi can achieve a performance of over 1TFLOP.
Intel Haswell (2013): Intel continues their push for lower power and more capable integrated graphics. While it didn't improve much over the previous generation, some new instructions were added and CPU voltage regulators were moved onto the processor package. The higher end GPU, dubbed Iris and Iris Pro, could go toe-to-toe with lower-end dedicated GPU offerings from AMD and NVIDIA. In addition, including the voltage regulators on the processor package allows fine tuning of power states for even more power saving. The only issue though is that presents a problem for overclockers which need to adjust CPU voltage at times as well as using thermal paste in the heatspreader. The latter issue resulted in Intel producing an intermediate revision, Devil's Canyon (which is clocked higher, uses better transfer material and has an improved electrical design) ahead of the core's major overhaul, Broadwell. The enthusiast version, Haswell-E upped the minimum core count from four cores to six, and also brought in the first consumer eight-core chip, while the Xeon version, Haswell-EP increased the maximum number of cores all the way up to eighteen. In the event, however, Broadwell only had a very limited lifespan on the desktop before Intel moved onto their next major product family, though the Broadwell-EP server chip would go on to have a much longer life.
AMD Kaveri (Steamroller) (2014): Heterogeneous System Architecture(HSA) comes to the x86-64 architecture. Previous PC APUs forces the memory controller to partition the main memory into GPU and CPU partitions (why these systems often report less RAM than installed in the BIOS screen- a portion of the RAM has been partitioned off from the CPU and reserved exclusively for the GPU). With HSA, both the CPU and GPU can work on a common memory and no partitioning is necessary. This theorically allows the GPU and CPU to use memory and communicate with each other much more efficiently. Allegedly, HSA has been used on ARM-based devices like smartphones for years, but this is the first time it would be implemented on PCs.
Berkely RISC-V (2014): Not an actual CPU product, but a new way that CPU instruction sets would be developed and licensed. As the name implies, it's a RISC-based ISA, but what sets it apart from others is that it's an open standard, meaning anyone can use the ISA and anyone can contribute. In addition, all of the software tools are open as well. This means that there's no hinderance as to what an instruction is supposed to do (or even how it might do it) in the RISC-V ISA. This is unlike x86, which Intel may say what an instruction does, but doesn't go into much detail in how it does it. In addition, any additional features may also be locked behind a license to get details on the specifies of the instruction.
Intel Skylake (2015): Intel's longest-lasting flagship design, though not really for the reasons they had intended. It launched as a decent, if unspectacular upgrade over Haswell, and by the end of its lifespan was a distant second place behind AMD in consumer performance,note and an even more distant fourth place behind AMD, IBM, and ARM in server performance. This time around, the problem wasn't so much with the CPU's design as Intel's 10 nm process for years being an utter disaster which was completely unable to produce a usable CPU, forcing Intel to abandon the core's planned 10 nm revision, Cannon Lake, in favor of the Coffee Lake and Comet Lake revisions, which bumped the core count up from 4 to 8 and 10 respectively, with no other major changes. This generation also largely saw the demise of the High-End Desktop CPU (in this case, Skylake-X) market, as the Zen 2 could fit the needs of people who needed large core counts in a normal desktop CPU package, while the other HEDT features such as more memory and PCI-E connectivity increasingly became seen as Awesome, but Impractical (it didn't help that multi-GPU set-ups all but died around this time as well).
AMD Zen (2016): After the failure of Bulldozer and its descendents, AMD rehires famed CPU architect Jim Keller note to work on the next architecture. Rather than try something novel or different, Keller's approach was to return to a traditional form with CPU design, this time incorporating simultaneous multithreading like Intel's HyperThreading to service up to two tasks per core and each core having its own floating point unit. Along with designing with efficiency and modularity in mind, AMD's first cut, released as Ryzen, Threadripper, and Epyc, wouldn't allow them to catch up to Intel quite yet, but was enough to get everyone's attention. Especially with Threadripper and Epyc, as instead of employing single-die designs like Intel, AMD decided to use multiple dies they called chiplets. This allowed them to effectively build whatever CPU they could want easily: Want more cores? Just add more chiplets!
The next iteration, dubbed Zen+ was more a refinement to tone down thermals and increase clock speed. It does however also corrects several errata with the design and introduced several new instructions that would allow the chip family to support Windows 11 officially.
The second generation, Zen 2, decoupled the I/O portion of the CPU into its own chiplet, allowing for even further modularity and was the first serious threat to Intel in over a decade, especially as Intel was still reeling in from its 10nm production woes. This allowed AMD to launch a then-unprecdented 64-core, 128-thread CPU. In addition, Zen 2 would also bring PCIe 4.0 to the masses, just as PCIe 4.0 SSDs with blazing fast speedsnote start appearing and the first generation of AMD's own Navi GPUs, which would also natively support PCIe 4.0, would hit the market. However, the memory controller for Zen 2 CPUs was extremely picky and required the timings of the DIMM modules to be even for maximum stability, an issue that would plague some users, and only be resolved in the next generation of Zen CPUs.
The third generation of Zen chips was when AMD had a solid advantage over Intel in performance. Late in the Zen 3's life, AMD also released a sort of experimental CPU with its touted 3D Cache (or stacking cache chips on top of each other), which led to even more performance in time sensitive tasks such as games.
Intel Ice Lake (2019): After nearly four years of trying fruitlessly to produce a 10 nm CPU, this was the first usable product Intel produced with the process. While it represented a decent improvement in performance per-clock over Skylake and its derivatives, the manufacturing issues limited Ice Lake to just quad-core models at low clock speeds. By all accounts this underwhelmed Apple so much that it motivated them into terminating their partnership with Intel (or at least sped up the inevitable), with Apple only bothering to use the chip in the Macbook Air. Intel later produced an upgraded version, Tiger Lake, which increased the clockspeeds, doubled the core count, and had a few other tweaks to the architecture.
- Processors in this family: Celeron, Pentium Dual-Core, Core i3, Core i5, Core i7
- Clock speed: 2.8 to 3.5 GHz (3.9 GHz with turbo)
- Core count: 2-8 (Sandy Bridge); 2-12 (Ivy Bridge)
- Word length: 64-bits (16 and 32 bits via hardware emulation)
- Address space: 64-bits (again, limited to 48-bits)
- Fab process: 32 nm (Sandy Bridge); 22 nm (Ivy Bridge)
- Used in: PCs
AMD Fusion (Bobcat, Llano, Jaguar) (2011): After AMD bought ATi in 2006, AMD had goals to merge the logical portion of the processor, with the computational powerhouse of the GPU. AMD eventually came out with a similar solution to that of Intel's in 2011 named Fusion. AMD's solution has more of a GPU focus, die shots showing that AMD's "Accelerated Processing Units" (APU) are roughly half GPU and half CPU. These features are rather popular with laptops, as graphical and computing performance increases dramatically, with little impact to battery life. In late 2012, the updated Trinity series of APUs was released, finally making the line viable for desktop users.
- Processors in this family: AMD A-series, E-series, embedded APUs for PlayStation 4 and X Box One (Liverpool and Durango).
- Clock speed: up to 4.2GHz
- Core count: 2-4, except Liverpool and Durango which has 8 each.
- Word length: 64-bits (16 and 32-bit via hardware emulation)
- Address space: 64-bits (again, limited to 48-bits)
- Used in: PCs, PlayStation 4, Platform/1XBoxOne
AMD FX-series (2011): AMD takes a radical approach to CPU organization in the processor. Instead of a "core", this processor has "modules", consisting of two basic execution units, with a shared FPU. There was claim that this used die space more efficiently. While the first version, "Bulldozer" did excel in multi-threaded tasks (up to the same performance as Intel's second generation Core i7 in some cases), in other tasks, it's only slightly better (or even slightly slower) than its predecessor. The second version, "Piledriver" improved matters — especially after AMD released the 9000-series with ludicrously high clock speeds — consistently equalling or beating the equivalent Core i7s in heavily threaded situations, albeit at the cost of much higher power consumption (about 2.5x that of Intel's offerings) and still-low single thread performance. However, AMD conceded that the line had no long-term future, and pulled the plug on FX versions of the next two planned cores, Steamroller and Excavator (though both found their way into their APUs). As of 2014, the series was repurposed as a part of the branding used on Fusion APUs, with the upcoming two FX chips being APUs instead of being discreet CPU-only chips.
- Clock speed: 3.2GHz to 4.7GHz (5.0GHz with turbo)
- Module/core count: 2-4/4-8
- Word length: 64-bits (16 and 32 bit via hardware emulation)
- Address space: 64-bits (again, limited to 48-bits)
- Used in: PCs
Second Generation Elbrus 2000 (2011)An updated e2k CPU with smaller tech norms and increased clock speed. Comparable to low-end Intel chips clocked several times faster in x86 emulation mode, and lightning-fast on the native VLIW code. Also extremely cool for its performance, radiating ~15-20W at full load. Recent versions also included several DSP cores to speed-up the image and signal processing, a vestige of its previous use as a radar CPU. While still mainly used in military applications, MCST has recently demonstrated a PC-standard mini-ITX motheboard, indicating their wish to enter the general PC market, especially multimedia servers, where their additional DSP cores would be especially useful.
- Clock speed: 500MHz to 1.2GHz
- Core count: 2-4 (CPU), 4-8 (optional DSP)
- Word length: 64 bit in native mode, 32 bit in x86 mode.
- Command word might be up to 2048 bit wide
- Address space: 64 bit
- Fab Process: 90 and 65 nm
- Used in: Military computers, PCs
Advanced RISC Machines ARMv8-A (2011): ARM goes 64-bit. With the demand for multitasking smartphones and tablets on the rise, and the demand for CPU-intensive gaming going mainstream, it is starting to become clear that the 17-year-old 32-bit ARM Cortex(ARM7TDMI-derived) instruction set will not be sufficient. This evolution promised that the ARM CPU will be able to handle more precise (and larger) numbers and address more memory, something important with large games.
- Processors in this family: NVidia's Tegra K1 Denver, Apple's A7/A8, Qualcomm's Snapdragon 410/808/810, Samsung Exynos 5433, and MediaTek MT6732 are among those that will most likely be the most widespread. Even AMD is releasing a 64-bit ARM CPU (Opteron A1100) to test the viability of using the ARM architecture in the Server market.
- Clock Speed: none defined. So far most CPUs go up to 2 GHz
- Core count: Up to 8
- Word length: 64 bits (compatible with ARM Cortex 32-bit instructions)
- Address space: up to 64 bits, part-specific. The Tegra K1 Denver, for example, caps out at 40 bit.
- Used in:
- iPhone 5s, iPhone 6/6 Plus, iPad Air, iPad Mini 2, Samsung Galaxy Note 4 (Asia/Europe version), HTC Nexus 9
Intel Xeon Phi (2012): The Xeon Phi started its roots in the ambitious GPU Larabee back in 2007, which Intel wanted it to do real-time ray tracing (the Holy Grail of real-time graphics, which in the end would not be achieved until nVidia released their Turing GPU in 2018). However, this fell through and was brought up on and off in 2010. The end result though, is what Intel dubbed the Many-Integrated Core (MIC) processor for highly parallel processing. Using a revamped P5 architecture (the same one in the original Pentium), the low-end Phi can achieve a performance of over 1TFLOP.
- Clock Speed: 1.053GHz
- Core count: 60
- Word length: 64-bits (16 and 32-bit via hardware emulation)
- Address space: 64-bits
- Used in: Supercomputers
Intel Haswell (2013): Intel continues their push for lower power and more capable integrated graphics. While it didn't improve much over the previous generation, some new instructions were added and CPU voltage regulators were moved onto the processor package. The higher end GPU, dubbed Iris and Iris Pro, could go toe-to-toe with lower-end dedicated GPU offerings from AMD and NVIDIA. In addition, including the voltage regulators on the processor package allows fine tuning of power states for even more power saving. The only issue though is that presents a problem for overclockers which need to adjust CPU voltage at times as well as using thermal paste in the heatspreader. The latter issue resulted in Intel producing an intermediate revision, Devil's Canyon (which is clocked higher, uses better transfer material and has an improved electrical design) ahead of the core's major overhaul, Broadwell. The enthusiast version, Haswell-E upped the minimum core count from four cores to six, and also brought in the first consumer eight-core chip, while the Xeon version, Haswell-EP increased the maximum number of cores all the way up to eighteen. In the event, however, Broadwell only had a very limited lifespan on the desktop before Intel moved onto their next major product family, though the Broadwell-EP server chip would go on to have a much longer life.
- Processors in this family: Celeron, Pentium, Core i3, Core i5, Core i7
- Clock speed: 2.4 to 4.0 GHz (4.4 GHz with turbo)
- Core count: 2-4 (Haswell, Devil's Canyon, and Broadwell); 6-18 (Haswell-E/EP); 6-22 (Broadwell-E/EP)
- Word length: 64-bits (16 and 32-bit via hardware emulation)
- Address space: 64-bits (again, limited to 48-bits)
- Fab process: 22 nm (Haswell); 14 nm (Broadwell)
- Used in: PCs
AMD Kaveri (Steamroller) (2014): Heterogeneous System Architecture(HSA) comes to the x86-64 architecture. Previous PC APUs forces the memory controller to partition the main memory into GPU and CPU partitions (why these systems often report less RAM than installed in the BIOS screen- a portion of the RAM has been partitioned off from the CPU and reserved exclusively for the GPU). With HSA, both the CPU and GPU can work on a common memory and no partitioning is necessary. This theorically allows the GPU and CPU to use memory and communicate with each other much more efficiently. Allegedly, HSA has been used on ARM-based devices like smartphones for years, but this is the first time it would be implemented on PCs.
- Processors in this family: AMD A-series, E-series, FX-series (FX-series is currently laptop-only).
- Clock speed: 1.35 GHz to 3.7 GHz (4.0 GHz with Turbo)
- Core count: 2/4
- Word length: 64-bits (16 and 32-bit via hardware emulation)
- Address space: 64-bits
- Fab process: 28 nm
- Used in: PCs, laptops.
Berkely RISC-V (2014): Not an actual CPU product, but a new way that CPU instruction sets would be developed and licensed. As the name implies, it's a RISC-based ISA, but what sets it apart from others is that it's an open standard, meaning anyone can use the ISA and anyone can contribute. In addition, all of the software tools are open as well. This means that there's no hinderance as to what an instruction is supposed to do (or even how it might do it) in the RISC-V ISA. This is unlike x86, which Intel may say what an instruction does, but doesn't go into much detail in how it does it. In addition, any additional features may also be locked behind a license to get details on the specifies of the instruction.
Likely since its inception, there are many CPU cores based on RISC-V, with many companies in later years shifting over to custom RISC-V cores of their own for controller chips or whatnot in place of existing designs from companies that have either long since abandoned them or the company no longer exists.
Intel Skylake (2015): Intel's longest-lasting flagship design, though not really for the reasons they had intended. It launched as a decent, if unspectacular upgrade over Haswell, and by the end of its lifespan was a distant second place behind AMD in consumer performance,note and an even more distant fourth place behind AMD, IBM, and ARM in server performance. This time around, the problem wasn't so much with the CPU's design as Intel's 10 nm process for years being an utter disaster which was completely unable to produce a usable CPU, forcing Intel to abandon the core's planned 10 nm revision, Cannon Lake, in favor of the Coffee Lake and Comet Lake revisions, which bumped the core count up from 4 to 8 and 10 respectively, with no other major changes. This generation also largely saw the demise of the High-End Desktop CPU (in this case, Skylake-X) market, as the Zen 2 could fit the needs of people who needed large core counts in a normal desktop CPU package, while the other HEDT features such as more memory and PCI-E connectivity increasingly became seen as Awesome, but Impractical (it didn't help that multi-GPU set-ups all but died around this time as well).
- Processors in this family: Celeron, Pentium, Core i3, Core i5, Core i7, Core i9
- Clock speed: 2.4 to 4.0 GHz (5.0 GHz with turbo)
- Core count: 2-4 (Skylake); 2-8 (Coffee Lake); 4-10 (Comet Lake); 4-28 (Skylake-EP)
- Word length: 64-bits (16 and 32-bit via hardware emulation)
- Address space: 64-bits (limited to 48-bits)
- Fab process: 14 nm
- Used in: PCs, laptops.
AMD Zen (2016): After the failure of Bulldozer and its descendents, AMD rehires famed CPU architect Jim Keller note to work on the next architecture. Rather than try something novel or different, Keller's approach was to return to a traditional form with CPU design, this time incorporating simultaneous multithreading like Intel's HyperThreading to service up to two tasks per core and each core having its own floating point unit. Along with designing with efficiency and modularity in mind, AMD's first cut, released as Ryzen, Threadripper, and Epyc, wouldn't allow them to catch up to Intel quite yet, but was enough to get everyone's attention. Especially with Threadripper and Epyc, as instead of employing single-die designs like Intel, AMD decided to use multiple dies they called chiplets. This allowed them to effectively build whatever CPU they could want easily: Want more cores? Just add more chiplets!
The next iteration, dubbed Zen+ was more a refinement to tone down thermals and increase clock speed. It does however also corrects several errata with the design and introduced several new instructions that would allow the chip family to support Windows 11 officially.
The second generation, Zen 2, decoupled the I/O portion of the CPU into its own chiplet, allowing for even further modularity and was the first serious threat to Intel in over a decade, especially as Intel was still reeling in from its 10nm production woes. This allowed AMD to launch a then-unprecdented 64-core, 128-thread CPU. In addition, Zen 2 would also bring PCIe 4.0 to the masses, just as PCIe 4.0 SSDs with blazing fast speedsnote start appearing and the first generation of AMD's own Navi GPUs, which would also natively support PCIe 4.0, would hit the market. However, the memory controller for Zen 2 CPUs was extremely picky and required the timings of the DIMM modules to be even for maximum stability, an issue that would plague some users, and only be resolved in the next generation of Zen CPUs.
The third generation of Zen chips was when AMD had a solid advantage over Intel in performance. Late in the Zen 3's life, AMD also released a sort of experimental CPU with its touted 3D Cache (or stacking cache chips on top of each other), which led to even more performance in time sensitive tasks such as games.
- Processors in this family: Ryzen, Threadripper, Epyc
- Clock speed: Up to 4.1 GHz (4.9 GHz with Turbo)
- Core count: Up to 64
- Word length: 64-bits
- Address space: 64-bits
- Fab process: 14nm (Zen), 12nm (Zen+), 7nm (Zen 2 & 3)
- Used in:
- PCs, laptops
- PlayStation 5
- Xbox Series X|S
- Atari VCS
- Steam Deck.
- Servers, Supercomputers.
Intel Ice Lake (2019): After nearly four years of trying fruitlessly to produce a 10 nm CPU, this was the first usable product Intel produced with the process. While it represented a decent improvement in performance per-clock over Skylake and its derivatives, the manufacturing issues limited Ice Lake to just quad-core models at low clock speeds. By all accounts this underwhelmed Apple so much that it motivated them into terminating their partnership with Intel (or at least sped up the inevitable), with Apple only bothering to use the chip in the Macbook Air. Intel later produced an upgraded version, Tiger Lake, which increased the clockspeeds, doubled the core count, and had a few other tweaks to the architecture.
- Processors in this family: Celeron, Pentium, Core i3, Core i5, Core i7
- Clock speed: 1.1 to 2.3 GHz (4.1 GHz with turbo)
- Core count: 2-4 (Ice Lake), 2-8 (Tiger Lake)
- Word length: 64-bits (16 and 32-bit via hardware emulation)
- Address space: 64-bits (limited to 48-bits)
- Fab process: 10 nm
- Used in: Laptops.
2020s
Apple Silicon (2021): Apple's first in-house CPU design for their iMac, MacBook, and the new Mac Studio line. Essentially, a performance oriented version of the CPU used in their phones and tablets. When first launched, various benchmarks pegged it at the highest performing CPU ever (in terms of the "instructions per clock" metric). In addition, it consumed so little energy that in testing, Apple thought there was a bug in their energy measurement software. Because of all this, Apple made a near immediate transition from using Intel's CPUs to their own, with no product refresh in sight for any of the Intel based products. Much like the CPUs in the mobile products, it also uses a hybrid design splitting the cores into high performance and high efficiency.
Apple would follow up the M1 with the M2 in 2022, with the main improvements being cores based on the A15 CPU (as opposed to the A14 cores used by the M1) alongside faster memory and a larger cache amount. They would then follow through again with the M3 in 2023.
Intel Rocket Lake AKA Intel 11th Gen (2021): Intel's fourth major blunder, and seen as a contender with Prescott for the title of their worst x86 processor. With Intel's 10 nm troubles seeming no nearer to being solved, they took the desperate move of trying to port the Ice Lake core over to the older 14 nm process in a desperate attempt to get out something that could compete with the Zen 2 and Zen 3 in performance-per-clock. However, Intel had to limit the speeds and core count (it actually has two fewer cores than the previous Comet Lake) of the chip in order to keep its thermals under control. The end result of this was that, outside of a few fringe cases, Rocket Lake often fared even worse against AMD's offerings than the Skylake derivatives were already doing. To add insult to injury, the new Alder Lake family was released later that year and finally restored Intel to competitiveness against AMD, leaving most enthusiasts confused as to why Intel had even bothered releasing this family.
Intel Alder Lake AKA Intel 12th Gen (2021): Intel's attempt at striking back against AMD's Zen, using a relatively new design in mainstream PCs desktops and laptops: hybrid CPU cores. This type of design adds performance cores (P-cores) for tasks that are time sensitive with efficient cores (E-cores) for tasks that don't need time sensitivity, such as background tasks. Also this CPU represents Intel finally being able to escape from its manufacturing woes, with this lineup being the first time their 10nm process was available across all platforms (Intel's 10nm was available in previous models, but were limited to mobile computers). While total performance did manage to beat AMD's Zen 3, it does consume a lot more power to do so.
AMD Zen 4 (2022): AMD trade blows with Intel and fights back against Alder Lake. Initially wanting to stay with PCIe 4.0, AMD quickly updated the CPUs to use PCIe 5.0 when it was announced that Intel's Alder Lake CPUs would natively support the new standard, and simultaneously a number of SSD manufacturers (most notably Corsair) pledging to release PCIe 5.0 SSDs that are capable of hitting read speeds as fast as 10GB/s alongside Intel's CPU. Aside from that, the CPU would also adopt DDR5 memory following Intel's footsteps. However, it will not adopt the hybrid CPU core system used by Intel- AMD's reasoning is that their CPU is so power efficient that it isn't necessary.
Apple would follow up the M1 with the M2 in 2022, with the main improvements being cores based on the A15 CPU (as opposed to the A14 cores used by the M1) alongside faster memory and a larger cache amount. They would then follow through again with the M3 in 2023.
- Processors in this family: M1, M1 Pro, M1 Max, M1 Ultra, M2, M2 Pro, M2 Max, M3, M3 Pro, M3 Max, M3 Ultra.
- Clock speed: Up to 3.2 GHz
- Core count: Up to 16 high performance cores, 4 high-efficiency cores
- Word length: 64-bits
- Address space: 64-bits
- Fab process: TSMC 5nm
- Used in:
- Debuted in a special version of the Mac Mini in 2020, to get developers started on porting apps to the new chip.
- All of Apple's computers from 2021 onwards
- iPad Pro 2021
Intel Rocket Lake AKA Intel 11th Gen (2021): Intel's fourth major blunder, and seen as a contender with Prescott for the title of their worst x86 processor. With Intel's 10 nm troubles seeming no nearer to being solved, they took the desperate move of trying to port the Ice Lake core over to the older 14 nm process in a desperate attempt to get out something that could compete with the Zen 2 and Zen 3 in performance-per-clock. However, Intel had to limit the speeds and core count (it actually has two fewer cores than the previous Comet Lake) of the chip in order to keep its thermals under control. The end result of this was that, outside of a few fringe cases, Rocket Lake often fared even worse against AMD's offerings than the Skylake derivatives were already doing. To add insult to injury, the new Alder Lake family was released later that year and finally restored Intel to competitiveness against AMD, leaving most enthusiasts confused as to why Intel had even bothered releasing this family.
- Processors in this family: Core i5, Core i7, Core i9
- Clock speed: 1.8 to 3.6 GHz (5.1 GHz with turbo)
- Core count: 6-8
- Word length: 64-bits (16 and 32-bit via hardware emulation)
- Address space: 64-bits (limited to 48-bits)
- Fab process: 14 nm
- Used in: PCs
Intel Alder Lake AKA Intel 12th Gen (2021): Intel's attempt at striking back against AMD's Zen, using a relatively new design in mainstream PCs desktops and laptops: hybrid CPU cores. This type of design adds performance cores (P-cores) for tasks that are time sensitive with efficient cores (E-cores) for tasks that don't need time sensitivity, such as background tasks. Also this CPU represents Intel finally being able to escape from its manufacturing woes, with this lineup being the first time their 10nm process was available across all platforms (Intel's 10nm was available in previous models, but were limited to mobile computers). While total performance did manage to beat AMD's Zen 3, it does consume a lot more power to do so.
Aside from that, Intel would become the first to adopt the PCIe 5.0 signal specification and DDR5 memory specification. However, the Alder Lake CPUs will retain backwards compatibility with DDR4 memory to ease transition. Intel later released a revision, Raptor Lake, which doubled the E-core count and made various other tweaks to the design, helping it stay competitive with the newly-released Zen 4.
- Processors in this family: Celeron, Pentium Gold, Core i3/i5/i7/i9
- Clock speed:
- P-cores: Up to 3.7 GHz (5.3 GHz with turbo)
- E-cores: Up to 2.8 GHz (3.6 GHz with turbo)
- Core count: Up to 8 P-cores and 8 E-cores (Alder Lake); up to 8 P-cores and 16 E-cores (Raptor Lake)
- Word length: 64-bits
- Address space: 64-bits
- Fab process: 10nm (dubbed Intel 7)
- Used in:
- PCs, laptops
AMD Zen 4 (2022): AMD trade blows with Intel and fights back against Alder Lake. Initially wanting to stay with PCIe 4.0, AMD quickly updated the CPUs to use PCIe 5.0 when it was announced that Intel's Alder Lake CPUs would natively support the new standard, and simultaneously a number of SSD manufacturers (most notably Corsair) pledging to release PCIe 5.0 SSDs that are capable of hitting read speeds as fast as 10GB/s alongside Intel's CPU. Aside from that, the CPU would also adopt DDR5 memory following Intel's footsteps. However, it will not adopt the hybrid CPU core system used by Intel- AMD's reasoning is that their CPU is so power efficient that it isn't necessary.
The thing that makes Zen 4 stands out is the departure from the six year old PGA type AM4 socket to a new LGA type socket coined AM5. AMD would also officially claim support for the USB4 standard in marketing materials, however it will be up to the motherboard manufacturers to implement USB4 support, and even then only a handful of motherboard manufacturers will have motherboards that support USB4 at launch. It is thought that this is due to the unfortunate timing of USB4v2 announced in early September, which is too close to the launch of the new CPU, and some motherboard manufacturer deciding that it would be better to wait for the newer standard which promises speeds twice as fast as first generation USB4.
Unfortunately for AMD however, it appears that the Four Is Death trope is at play, as this CPU too is shrouded in controversy and problems.
Nearing launch, it was found that these CPUs would routinely run to 95C under synthetic loads- to the point where it is suggested that the CPUs be used with active water cooling. However, AMD insisted this is by design as the CPU boosts until it reaches its thermal limit first before dialing down clock speeds. This is unlike older designs that would stop boosting at a certain power limit first. Another valid concern raised by users was that unlike Intel, AMD was immediately dropping support for DDR4 with the CPU, choosing only to support DDR5. However, the price of DDR5 quickly fell in several months, and as of February 2023, is comparable to the price of DDR4 RAM.
Even later, it was discovered that some motherboards, particularly those from Asus, were aggressively driving the CPUs by overvolting them to the point that they were severely overheating and prematurely failing (with the CPU bulging as a visual side-effect). As of May 2023, investigations concluded that the problem was twofold- firstly, AMD's microcode imposed no restrictions as to how much users are allowed to overclock their CPU, meaning that users who push their luck were basically allowed to tell the motherboard to murder the CPU. The second issue was that Asus' motherboards were already overclocking the CPU out of the box and then lying about the voltage fed to the CPU, meaning those who tried to push the CPU were in fact pushing their luck further. AMD's response was to lock the CPU's maximum voltage on their updated microcode to a safe level while still allowing undervolting, at the cost of the stability of memory overclocking and making their EXPO technology moot. However, Asus basically put themselves in a bad light by warning people those who used the updated BIOS that the new microcode comes with will be voiding their warranty. Even worse was that it was found out that Asus' BIOS was bugged and the new microcode didn't really work. Backlash was swift and Asus quickly backtracked on their warranty policy. This entire guffaw also made AMD announce that they will be sunsetting their current AGESA microcode for the OpenSIL open-source microcode solution.
About 3 months after the launch of the Zen 4 CPUs, AMD would launch the X3D variants of the chips. These chips would pack three times the amount of cache compared to the normal variants. Of the three chips launched, the 7800X3D would become the desirable option as the 7950X3D and 7900X3D, while powerful, was designed such that only 8 of the 16 cores had 3D cache. This resulted in the need to park the non-3D-cache cores when performance is needed, as the performance will be worse if those cores were used in conjunction with the 3D cache cores. On the other hand the 7800X3D had no such issues since it all of the cores had access to the 3D Cache.
This was shortly after followed with the Ryzen 8000 series of APUs, which would drop the chiplet design for a monolithic design once more, but include a relatively powerful GPU unit on-die. At the same time, hints of the Ryzen 9000 series of CPUs dropped, with a leaked E-mail from HP's engineering team to AMD's in regards to the CPU.
- Processors in this family: Ryzen 5/7/9 7000/8000/9000 series
- Clock Speed: Base at 4.5 GHz (Up to 5.7 GHz with Turbo)
- Core count: Up to 16 cores
- Word length: 64-bits
- Address space: 64-bits
- Fab process: 5nm
- Used in: PCs
