Detailed · 24 events
A History of Semiconductors and Hardware
1940s
John von Neumann / Internet Archive scan of Samuel N. Alexander's copy (Wikimedia Commons) · Public Domain · Commons ↗ At the Moore School of Electrical Engineering, University of Pennsylvania, an unfinished typescript titled First Draft of a Report on the EDVAC went out bearing John von Neumann's name alone. It divided a computing machine into a central arithmetic part (CA), a central control (CC), a memory (M), input (I) and output (O), and described a store holding instructions and numbers alike — all in the language of logic, stripped of vacuum tubes and delay lines. Being a draft, its forward references were left blank, and it cited exactly one paper: McCulloch and Pitts on neural nets. Eckert and Mauchly disputed the attribution, and the unrestricted circulation was later held in court to be a prior public disclosure that voided their patent claim.
U.S. Army Photo (Wikimedia Commons) · Public Domain · Commons ↗ ENIAC, built for the Army Ordnance Department's ballistics calculations, was shown to the press at the Moore School of Electrical Engineering, University of Pennsylvania; the formal dedication followed on 15 February. It held 17,468 vacuum tubes, weighed about 30 tons, and added roughly 5,000 numbers a second. It was not, at its unveiling, a stored-program machine: each problem was set up by replugging cables and turning switches. The demonstration's trajectory program was written by Betty Snyder and Jean Jennings, and the six women who programmed the machine are almost entirely absent from the coverage. Nor was it the first electronic computer — Colossus and the Atanasoff-Berry Computer came earlier. What ENIAC was first at is being a general-purpose programmable electronic computer that was actually finished, actually used, and publicly known.
1950s
U.S. Census Bureau (via Wikimedia Commons) · Public domain · Commons ↗ The US Census Bureau accepted the first UNIVAC I on 31 March 1951. It was designed by ENIAC's J. Presper Eckert and John Mauchly, whose Eckert-Mauchly Computer Corporation had run out of money and been sold to Remington Rand in February 1950, so UNIVAC I reached the world as an acquirer's product. It was not the world's first commercial computer: the Ferranti Mark 1 was delivered to the University of Manchester in February 1951. What UNIVAC I was first at is narrower — the first commercial electronic computer built in the United States, and the first computer received by a civilian US government agency. Serial number 1 stayed at the Philadelphia plant after acceptance and was not moved into a Census Bureau building until December 1952. On election night, 4 November 1952, a UNIVAC appeared on CBS television and produced an early forecast of an Eisenhower landslide that was never put on the air.
Arnold Reinhold (Wikimedia Commons user ArnoldReinhold) · CC BY-SA 3.0 · Commons ↗ IBM announced the 305 RAMAC at a press conference. Its central component, the IBM 350 disk storage unit, was the first commercial hard disk drive. A stack of 24-inch disks presented 100 recording surfaces, spun at 1,200 rpm, and a single pair of heads moving up and down and in and out reached any record in about 0.6 seconds on average. Capacity was five million characters — 3.75 MB if you count six data bits per character, 5 MB if you count a character as a byte, a number whose answer depends on the conversion. Tape forced you to read down the reel until you reached the record you wanted; RAMAC took the same time to reach any of them. That is where looking up stock or an account on the spot, rather than in the overnight batch, begins. The 305 system leased for US$3,200 a month.
Florian Schäffer (Wikimedia Commons user MAbW) · CC BY-SA 4.0 · Commons ↗ Jack Kilby at Texas Instruments etched transistor, capacitor, and resistor elements into a single bar of germanium 7/16 by 1/16 inch (11.1 by 1.6 mm), joined them with fine gold flying wires, and ran the resulting phase-shift oscillator in front of his managers. It was the first working integrated circuit. On 23 January 1959, Robert Noyce at Fairchild Semiconductor independently conceived a monolithic IC on silicon that integrated the interconnect as well, building on Jean Hoerni's planar process. Kilby filed his patent on 6 February 1959 and Noyce on 30 July 1959; it was Noyce's planar form that became the basis of volume manufacturing. Kilby received the 2000 Nobel Prize in Physics, shared with Zhores Alferov and Herbert Kroemer.
1960s
ArnoldReinhold (Wikimedia Commons) · CC BY-SA 3.0 · Commons ↗ IBM announced System/360, a family of computers spanning the whole price and performance range on one architecture. The Data Processing Division press release issued that day records that more than 100,000 people attended announcement meetings in 165 American cities, that the central processors offered 19 combinations of speed and memory capacity, and that more than 40 types of peripheral equipment came with them. The product on sale was not speed but compatibility: a program written for one model ran on any other in the line with the same class of I/O, ending the practice of throwing away software with every upgrade. IBM's own account puts development at USD 5 billion over four years — more than the company's annual revenue at the time. Deliveries were announced for the third quarter of 1965 (small configurations) and the first quarter of 1966 (largest); of the six models announced, the 60, 62 and 70 never shipped and were replaced by the 65 and 75. The struggle to deliver OS/360 produced Fred Brooks's The Mythical Man-Month.
Max Roser, Hannah Ritchie / Our World in Data (Wikimedia Commons) · CC BY 4.0 · Commons ↗ Gordon Moore, head of R&D at Fairchild Semiconductor, contributed an article to Electronics magazine projecting that the number of components on a chip would double roughly every twelve months—an extrapolation from the 1959-64 data he had. He revised it to a doubling every two years at an IEEE meeting in 1975, and it is that revised figure, not the 1965 one, that circulates today as Moore's Law and became the semiconductor industry's de facto roadmap. In 1968 Moore and his colleague Robert Noyce co-founded Intel.
1970s
the Science Museum (Science Museum Group), via Wikimedia Commons · CC BY 4.0 · Commons ↗ A 4-bit processor conceived by Ted Hoff and Stan Mazor at Intel, laid out by Federico Faggin in silicon-gate MOS, and finished with Masatoshi Shima, seconded from the Japanese calculator maker Busicom. 2,300 transistors on roughly 12 mm² in a 10 μm process, clocked at 740 kHz; an instruction cycle took 10.8 μs, about 93,000 instructions per second. It first shipped to Busicom in March 1971 for the 141-PF calculator and was announced to the public on 15 November 1971 in an Electronic News advertisement. It was the first single-chip CPU anyone could buy — earlier single-chip designs such as Four-Phase's AL1 and the F-14's MP944 were never sold as general-purpose parts. The 8-bit 8008 was not derived from the 4004 but was a separate design for Computer Terminal Corporation; from it came the 8080, the 8086, and x86.
The wub (Wikimedia Commons) · CC BY-SA 4.0 · Commons ↗ An experimental machine designed at Xerox's Palo Alto Research Center (PARC) in early 1973. It stood a bitmapped CRT on its end so the screen matched a sheet of paper, gave it a three-button mouse and a five-finger chord keyset, and wired it to an Ethernet — nearly every component of the modern workstation, assembled at once. Bravo made document editing look on paper the way it looked on screen; Smalltalk made object-oriented computing a live environment. The Alto was never sold as a product, yet by the summer of 1979 close to a thousand of them were in daily use by researchers, engineers and secretaries. The demonstrations Apple received here in December 1979 fed into the Lisa and the Macintosh.
The wub (Wikimedia Commons) · CC BY-SA 4.0 · Commons ↗ The Apple II was shown at the first West Coast Computer Faire, held 15-17 April 1977 at the San Francisco Civic Auditorium and Brooks Hall. It worked without opening the case, drove colour on a domestic television, carried BASIC in ROM, and had eight expansion slots. It shipped on 10 June 1977 at US$1,298 with 4 KB of memory. Steve Wozniak's design acquired colour and sound because he wanted to write Breakout in software rather than in hardware. The Disk II of 1978 got it off cassette tape, and VisiCalc in October 1979 gave it a reason to be bought by businesses — which is how a hobbyist's machine became office equipment.
1980s
The wub (Wikimedia Commons) · CC BY-SA 4.0 · Commons ↗ IBM announced its smallest, lowest-priced computer system: the press release put it at "as little as $1,565", with 16,384 bytes of standard memory, an 83-key keyboard, and first deliveries scheduled for October. The specification mattered more than the machine. The Boca Raton team bought off-the-shelf parts, published the circuit designs in a technical reference, and left Microsoft free to sell the same operating system to anyone else. The one piece IBM kept proprietary was the BIOS in ROM — and when Compaq, founded in 1982, reimplemented it by clean-room method, the compatible-machine industry had its opening. IBM's own corporate history quotes the historian James Cortada: the company's PC market share fell from roughly 80% in 1982–83 to 20% a decade later. In 2005 IBM sold the PC division to Lenovo.
Peter Howkins · CC BY-SA 3.0 · Commons ↗ At Acorn Computers in Cambridge, UK, the first silicon of ARM1—a RISC (Reduced Instruction Set Computing) processor designed by Sophie Wilson and Steve Furber—arrived and worked correctly the moment it was powered up, on 26 April 1985. The name then stood for Acorn RISC Machine. Just 25,000 transistors on a three-micron process: the smallness is what produced the low power draw that would define the architecture. ARM1 itself was never a product chip, though—it lived on the ARM Evaluation System, a second-processor board for the BBC Micro. The chip that actually reached a home computer was its successor, the ARM2, in the Acorn Archimedes of June 1987. In November 1990 the design team was spun out as Advanced RISC Machines Ltd., a joint venture of Acorn, Apple, and VLSI Technology, and ARM went on to become the default for embedded and mobile devices. Arm reports that more than 350 billion chips built on its architecture have now shipped.
Questions this page answers
- Which computer first shipped with an ARM chip?
- The Acorn Archimedes of June 1987, and it carried the ARM2, not the ARM1. ARM1 was never a product chip: it lived on the ARM Evaluation System, a second-processor board for the BBC Micro.
- What did ARM originally stand for?
- Acorn RISC Machine. The design team was spun out in November 1990 as Advanced RISC Machines Ltd., a joint venture of Acorn, Apple and VLSI Technology.
2000s
Hyins (Wikimedia Commons) · Public domain · Commons ↗ NVIDIA shipped CUDA 1.0 (Compute Unified Device Architecture), described in its own Programming Guide as a hardware and software architecture for computing on the GPU 'without the need of mapping them to a graphics API'. The idea arrived with the G80 (GeForce 8800) in November 2006; a public beta followed in February 2007; the 1.0 Programming Guide is dated 23 June 2007. Supported hardware was the GeForce 8 Series, Quadro FX 5600/4600, and Tesla. Scientific computing was the intended market, but the fit with deep learning demonstrated five years later by AlexNet (2012, trained on two GTX 580s through CUDA) is what turned NVIDIA into the central infrastructure company of AI computation.
2010s
- EVT.016T1AlexNet — The Deep-Learning Era BeginsA History of Artificial IntelligenceA General History of Information TechnologyRead more →
- EVT.017T2iPhone 5s — Touch ID and the 64-bit A7A History of the iPhoneA History of Mobile Phones and Smartphones
曾成訓 (Wikimedia Commons) · CC BY 2.0 · Commons ↗ Taiwan Semiconductor Manufacturing Company (TSMC) began volume production of its 7 nm process (N7). '7 nm' is a node name, not a measurable dimension on the die. Within the year 7 nm accounted for 9% of TSMC's wafer revenue, and N7+, which uses EUV lithography, moved from R&D into manufacturing (TSMC's 2018 Form 20-F). The first major part shipped in volume on the node was Apple's A12 Bionic in the iPhone XS of September 2018, which Apple billed as the first 7-nanometer chip in a smartphone. While Intel struggled to move to 10 nm, TSMC consolidated a dominant position in leading-edge logic manufacturing and came to fabricate much of the advanced silicon of Apple, AMD, NVIDIA and Qualcomm.
Questions this page answers
- Does 7 nm describe an actual measurement on the chip?
- No. It is a node name, not a measurable dimension on the die.
2020s
Sonic8400 · CC BY-SA 4.0 · Commons ↗ Apple announced M1—an in-house ARM-based system on a chip—and the first three Macs to carry it: MacBook Air, Mac mini, and the 13-inch MacBook Pro. Sixteen billion transistors on a 5-nanometre process integrated an eight-core CPU (four performance, four efficiency), a GPU of up to eight cores, and a 16-core Neural Engine, with the memory sitting in the same package. Apple claimed up to 3.5x the CPU performance, 6x the graphics, 15x the machine learning, and twice the battery life of previous-generation Macs—figures from Apple's own testing. This did not end the Intel era that began in 2006: Apple said plainly it still had 'exciting new Intel-based Macs in development,' and the transition only closed in June 2023 when the Mac Pro went to Apple silicon. It was, even so, the moment the decade-plus of in-house chip design that started with the iPhone and iPad finally reached the Mac.
Questions this page answers
- Did the M1 end the Intel Mac era?
- No. Apple said plainly at the time that it still had new Intel-based Macs in development, and the transition closed only in June 2023, when the Mac Pro moved to Apple silicon.
- Where does the 3.5x faster CPU figure come from?
- From Apple's own testing, against previous-generation Macs. The claims of up to 6x graphics, 15x machine learning and twice the battery life share that source; none is an independent measurement.
极客湾Geekerwan (Wikimedia Commons) · CC BY 3.0 · Commons ↗ At GTC on 22 March 2022 NVIDIA announced the Hopper architecture and its first GPU, the H100: 80 billion transistors on TSMC's 4N process, the first GPU to support PCIe Gen5 and to use HBM3 (NVIDIA's claim), 3 TB/s of memory bandwidth, and a Transformer Engine working in FP8. The SXM module carries 80 GB of HBM3; the PCIe card carries 80 GB of HBM2e. The headline speedups are NVIDIA's own, stated against 'the previous generation' — the Ampere-based A100 — with no precision given: 'up to 9x faster' training on a 395-billion-parameter Mixture of Experts model, and 'up to 30x higher throughput' for Megatron 530B chatbot inference. Full production was announced on 20 September 2022 and partner systems shipped from October. The demand surge that followed ChatGPT's launch that November made the H100 the de facto standard training and inference processor of the generative-AI era. The line continued through H200 (2023), Blackwell B200 (2024) and Blackwell Ultra (2025); on 31 May 2026 NVIDIA announced that the successor platform, Vera Rubin, had ramped into full production.
Daniel Torok / The White House (Wikimedia Commons) · Public domain (US Federal Government work) · Commons ↗ At GTC 2024, NVIDIA unveiled the Blackwell GPU architecture and its first product, the B200—an announcement, not a shipment. 208 billion transistors; two dies on TSMC's custom 4NP process joined by a 10 TB/s chip-to-chip link and presented as one logical GPU. NVIDIA's own figures: 20 PFLOPS of FP4 per GPU with 2:4 sparsity (10 PFLOPS dense), and 192 GB of HBM3e as specified at launch. The rack-scale GB200 NVL72 (36 Grace CPUs, 72 Blackwell GPUs) was claimed at 1.4 exaflops, with up to 30× the LLM-inference performance of the same number of H100s and up to 25× lower cost and energy. Yield problems in the CoWoS-L package pushed mass production to December 2024, but Blackwell became the spine of generative-AI compute infrastructure.
NVIDIA Corporation (Wikimedia Commons) · Apache License 2.0 (trademark applies) · Commons ↗ On 18 June, NVIDIA's market capitalisation reached about US$3.34 trillion, briefly overtaking Microsoft and Apple to become the world's most valuable company. A market acknowledgment of its transformation from gaming-GPU vendor into the central infrastructure company of AI compute. At the end of 2022, just after ChatGPT's debut, NVIDIA had been valued at roughly US$364 billion—a ninefold growth in under eighteen months. It reclaimed the top spot for good on 5 November 2024, closing above Apple, and Jensen Huang became one of the most influential executives in the AI industry.
