MIPS for the masses
Acorn Computers and ARM
Chris Curry and Herman Hauser worked with Clive Sinclair on a microcomputer kit around 1978, and the two had some ideas of their own, and they then left Sinclair and founded Cambridge Processor Unit Ltd on the 5th of December in 1978. Very shortly after they started CPU Limited, they hired Sophie Wilson, Steve Furber, and Chris Turner. The group then worked at building a computer. Sophie had a small system working fairly quickly, and wrote a system monitor for it (by hand in binary) that would allow programming in hex. After that came an assembler written with that monitor. Then came a BASIC interpreter written in that assembler.
In January of 1979, CPU Ltd began using the trade name of Acorn Computer Ltd. The Acorn System 1 was released released in March of 1979. This was an 8 bit microcomputer based around the MOS 6502 at 1 MHz and 1152 bytes of RAM. The computer was based around two eurocards. The first had CPU, RAM, ROM, support ICs, and the other had a 7 segment LED display, a 25 key keypad (hex and functions), and a cassette interface. The two cards were connected with a flexible cable.
Abort, Retry, Fail is a reader-supported publication. To receive new posts and support my work, consider becoming a free or paid subscriber.
The Acorn System 2 of 1980 improved upon this quite a bit. The change allowing the improvements started with cards going into an 8 slot backplane instead of being connected by a cable. The base RAM in the System 2 was expanded to 2K, the System 2 featured a proper keyboard interface. There was a VDU card allowing 40x25 text display, and there was a memory board with 4K RAM. This memory board also held a 4K ROM with BASIC. The memory could be further expanded by an additional 4K RAM and another 4K ROM that had floating point functions for the BASIC interpreter. If you do your addition, that puts the System 2 at 10K RAM with 4K ROM. This system was mounted in a 19 inch sub-rack frame. This provided the basis for systems 3, 4, and 5 which would add floppy controllers, more RAM, more expansion slots, a 2 MHz 6502, a network adapter, and an 80 character width display controller. The finaly year of shipment for Acorn rack-mounted systems was 1983.
The Acorn Atom
The Atom was, essentially, an Acorn System 3 shoved inside of a keyboard housing. It featured a MOS 6502 at 1 MHz, 2K base RAM (expandable to 12K), 8K ROM featuring BASIC and an assembler which could be expanded to 12K with optional add-on ROMs, single channel sound via built-in speaker, tape interface (KCS), TV out, and a Centronics connector. Some machines also featured an Econet adapter (ring network), and a PAL color card.
In the machine’s lowest configuration, its price would have been around £170 (£1016.52 in December 2022), and the highest configuration would have been around £250 (£1494.88 or 1836.16 USD in 2022).
Eight of these glorious microcomputers were demonstrated with networking in March of 1980. This really stands out to me due to how advanced this was for home micros at the time. The network allowed screen sharing, file sharing, and remote keyboard. That’s quite nifty.
Comparing the Atom to the 1979 ZX80 from Sinclair… the Atom is the easy winner, but the Atom was far more expensive. It was the development and sales of the Acorn Atom that pushed Chris Curry to properly incorporate Acorn Computers (instead of as a trade name of CPU Ltd) and begin working more seriously on growing the company. According to Sophie Wilson, Chris Curry had always wanted to make something better for the home micro user. Chris Curry ceased other work and shifted fulltime to Acorn.
The BBC Micro
The BBC (British Broadcasting Company) was founded just over 100 years ago (18th of October of 1922). It became the British Broadcasting Corporation 5 years later. The BBC first focused on radio, with their first broadcast made at 18:00 GMT on the 14th of November in 1922. Arthur Burrows read the news. Being the UK, the weather was absolutely in the very first broadcast. With such a humble start, the BBC rapidly expanded their library of programmes. By the 30s, all of the radio show genres were represented, and in ‘32 the BBC World Service was launched for British territories abroad (the Empire was still a thing). The BBC started TV broadcasts in ‘36 but took a hiatus for the War starting in ‘39, and TV broadcasts came back in ‘46.
The coronation of Queen Elizabeth II on the 2nd of June 1953 changed the direction of the BBC entirely (or can be used as a landmark in time when the change was made visibly certain). Up to this, radio had been dominant. After this, the dominance of TV was certain. By the end of the 50s, many programmes that had been broadcast on radio and TV had become TV only.
According the BBC’s website on the matter, in the late 1970s the UK felt as if it were falling behind in technological prowess. To address this, the BBC launched the Computer Literacy Project. The CLP involved several TV series, a BASIC programming course, learning materials, and computer software. All of the computer software (including the BASIC dialect used in educational materials) was for the BBC Micro (aka the “beeb”).
During the planning of the CLP, the BBC figured it would need a computer that had a lot of capability due to the topics the CLP would cover: programming, graphics, music, teletext, controlling physical hardware, networking, and some AI stuff. This effectively meant that only one computer of the time would be suitable, the Acorn Proton, which was the successor the Atom. Sinclair, Newbury, Tangerine, and Dragon had all tried to push their own machines for the project, but they all fell short in some capacity. As such, the Acorn Proton became the BBC Micro.
The BBC Micro was used by most schools in the UK, and over 2 million of them were sold. This machine propelled Acorn to fame, wealth, and success. What was the BBC Micro?
MOS 6502 at 2 MHz
cassette tape storage
floppy (5.25 and 3.5)
winchester HDD (rare)
16K standard RAM (models ranged all the way to 128K)
32K to 128K rom (optional roms can push this to 272K)
Texas Instruments SN76489 for sound, also a speech synthesizer
keyboard, joysticks, and lightpen for input
parallel, serial, econet
Tube interface for a second processor
The Tube interface is particularly interesting to me. When used, the 6502 processor in the BBC Micro was used for all I/O, and applications would run on the second processor. If one were using a second 6502, this meant he/she would get better performance with all of the beeb’s normal software. However, that wasn’t the only option. A beeb could use a Z80 with the Tube, and then run CP/M and all of the software from that ecosystem. The NS32016 could also be used and this could run Acorn’s Panos, or a UNIX variant.
All of this hardware wasn’t cheap. Released on the 1st of December in 1981, the price was £235 for an entry level unit (£1064.61 or 1307.66 USD in 2022). That base unit had no floppy drives, no hard disk, only 16K RAM, only 32K ROM, no networking, no joysticks, no lightpen. But, this system would give you the ability to expand over time.
Despite selling well, there was still more of a computer market to be had. Due to the cost of the beeb, Sinclair computers were selling quite well as they were much lower priced. So as a response to Sinclair, in 1983, Acorn released the Electron for £199. This was competitive given that the Electron was still more capable than the competition. This machine was 32K RAM, 32K ROM, a 6502 at 2MHz, and BBC BASIC II. This machine could use casettes, floppies, or ROM carts assuming that you purchased a cassete player, floppy drive, or ROM cart. For video output, the Electron had an RF modulator, RGB, or composite. Sound wasn’t great. The Electron had a single channel which amounted to beeps most of the time. Unfortunately for Acorn, the Electron really didn’t do well in the market. It did well enough to get software ports and expansions, but not well enough to encourage Acorn to continue working on the Electron. Despite not being the second huge commercial success that Acorn was hoping for, the machine wasn’t completely without merit. The ROM cartridge system and the improved disk filesystem (ADFS) that debuted on the Electron were later used in the BBC Master (the beefiest BBC Micro) and the Acorn Archimedes.
Acorn was doing very well in the early 80s with the BBC Micro and its variants. It was clear, however, that the market for personal computers was changing rapidly. The IBM PC had debuted in 1981, and by the middle of the 1980s, the IBM PC was becoming the dominant platform for both business and home computers. IBM compatible machines were showing up everywhere. The PC brought with it the 16 bit era. Additionally, the Apple Lisa launched in 1983, the Apple Macintosh launched in 1984, and the Commodore Amiga launched in 1985. These machines proved that the GUI was the future of home micros. Luckily, Sophie Wilson at ARM had somewhat anticipated this. According to her, the whole purpose of the tube was to bridge the gap between a cheap home micro and a proper workstation. She had done the bulk of the work on the BBC Micro’s tube interface, and this came in handy when the 16 bit machines started taking over the market. Acorn started testing multiple processors: 80286, 6809, NS32016, 68020. None of these were what they wanted.
Specifically, Sophie notes in an interview:
These high-end processors with their clock rates twice the 6502s with their bus widths twice the 6502s were actually, in many circumstances, slower. This was because they didn't use the memory system effectively. On the 6502 at 4 megahertz, each cycle is a memory access. So it accesses 4 megabytes a second. On all these other ones, a memory access is four cycles of the master clock on a 68000. So an 8 megahertz 68000 would only do two bus accesses per microsecond, and they would be two bytes wide. So they only had four megabytes bandwidth, just the same as the 6502. But in building the 6502-based systems, we knew how to make very high-performance memory systems. And if we made the memory system wider, a 16 or 32 bit wide memory system, we could easily design a memory system with 16 to 32 megabyte-per-second memory bandwidth. But we couldn't find a microprocessor that could eat that. We became convinced, which we think is still true for cacheless microprocessors, that memory system bandwidth is a direct predictor of system performance.
They were neither fast enough nor simple enough for the team at Acorn. They already had some custom IC design experience, but they weren’t too certain about designing a CPU from scratch. The visited both National Semiconductor and Western Design Center, and didn’t really like the CPUs that were being built by either company. Roughly around this time, the very first RISC designs were being published and made known by Berkley, Stanford, and IBM. IBM had already built the 801 and ROMP, but this wasn’t visible outside of IBM until this time. Herman Hauser made sure that some of the published IBM research was available to the team.
The Birth of ARM
Sophie Wilson, Steve Furber, and Hermann Hauser comprised the Acorn Advanced Research & Development team (AR&D) in the beginning. The three of them started working on designing the new CPU: the Acorn RISC Machine 1 (ARM). This was October of 1983. Sophie was working on the instruction set, Steve was working on some fundamental chip design work, and Hermann was helping to clarify ideas and provide some feedback. Once they got going, the team quickly expanded. Now, there wasn’t just a CPU, but an entire suite of chips: I/O controller, video controller, memory controller, the CPU. The team kept expecting to find a roadblock. They expected to quickly learn why no one developed chips outside of megacorps, but nothing seemed to stand in their way. Accordingly, the team put a lot of effort into modeling, into verification, and into testing. The initial modeling and testing was effectively a VM running on a 6502 coprocessor on the BBC Micro. This did exceptionally well. So well that by September of 1984 the design that the team had made was handed off to VLSI for production. This was a pure 32 bit design with 45 instructions, most of those executing in just one cycle, and this was enabled by 25000 transistors on a 3 micron process at 6 MHz. On the 26th of April in 1985, the first ARM CPU came back to AR&D. The ARM1 was slotted into the tube and the entire instruction set was tested. It ran perfectly. Amazingly, one time during testing, the chip didn’t power down. The power supply hadn’t been connected to the chip. The entire thing was running off of power leakage from the I/O pins. Acorn had been aiming for max power draw of one watt, but the actual power requirement found during testing was about one tenth of one watt. Despite having very low power requirements, the chip performed quite well in benchmarks. At the same clock speed, the ARM1 was about 10x the performance of the Intel 80286. They continued testing until July of 1985.
Initially, ARM was sold as a tube processor the BBC Micro. The CPU work was started before the support chips, and was therefore ahead by over a year. The rest of the chips came back, and they too functioned without error. These were then assembled into the ARM Evaluation System. Universities and researchers were happy to have more compute power, and they not only bought the CPU but also ported software to it. Acorn had already ported Prolog and LISP to the ARM, but unversity folks ported C and BCPL. An ecosystem was already forming.
The problem was that this was a bit late.
BBC Micros weren’t selling well by this point. The Sinclair Spectrum and newly introduced Commodore 64 (1983) were taking the low end of the market, and the IBM PC was taking the high end. The Amiga and the Macintosh were pushing the boundaries of computing forward. Acorn was losing ground very quickly. As I’ve written before, better products often fail. At this point if Acorn didn’t do something quickly, it would cease to exist. It isn’t as if Acorn was the only company struggling by the middle of the 1980s. Atari was split and sold two different directions. Despite being groundbreaking, poor performance of the Lisa and Macintosh meant troubles at Apple that ultimately caused Jobs to leave. There had been a recession in the late 70s and early 80s, and IBM entered the market when its competitors were at their weakest.
By 1985, Acorn was in serious debt and their share prices had collapsed. Chris Curry and Herman Hauser signed a deal with Olivetti on the 20th of February that reduced their ownership from roughly four fifths to just over a third of the company. Olivetti now owned roughly half of the company. This eased the financial pains of Acorn temporarily, but it wasn’t enough. In July, Olivetti purchased more of the company, raising their control to four fifths. Acorn was able to negotiate some debt forgiveness, and they managed to get the BBC to waive half of the royalty payments owed.
Through all of this turbulence, development of ARM continued. In 1986, the ARM2 was introduced. This chip had faster interrupts, a clock speed of 8MHz, and it was cheaper than its rivals while being more performant with 4 MIPS. The transistor count bumped to 30k. The following year, a 12 MHz ARM2 was released which delivered 7 MIPS. These were first sold as second CPUs to the newly released BBC Master (a beefier BBC Micro) that was sold to the educational market in the UK. Roughly 200000 of these were sold.
By the middle of June of 1987, Acorn had released the Archimedes. This was roughly a year and a half after the IBM RT. RISC was in the air.
True to the slogan of the development team “MIPS for the masses” and accomplishing their goal, the Archimedes delivered 4 MIPS at launch. The Archimedes came in four models at launch: the A305, the A310, the A410 and the A440. A little after the initial launch, the A540 was released and this model featured the ARM3 CPU which had an integrated cache of 4K, delivered 12 MIPS and ran at 25MHz. The ARM3 was also offered as an upgrade for the 300 and 400 machines. The Archimedes featured:
800KB 3.5” FDD (backward compatible with 640K disks)
102 key AT style keyboard, 3 button mouse
dedicated memory controller IC, I/O controller IC, video controller IC
21 video modes, among them were 640x480 (256 colours), 800x600 (16 colours), 1280x960 (monochrome), text - 132 x 32 maximum
8 voice stereo sound
external ports: Centronics, RS423, Video composite, RGB, Econet
ARM CPU with RAM up to 16MB
A305 - 512 K RAM - £799 (£2081.63 - 2538.24 USD in 2022)
A310 - 1 M RAM - £875 (£2279.63 - 2779.67 USD in 2022)
A410 - 1 M RAM - £1399 (£3644.80 - 4444.29 USD in 2022)
A440 - 4 M RAM, 40 M HDD - £2299 (£5989.56 - 7303.37 USD in 2022)
A540 - 16 M RAM, 120 M HDD - £2499 (£6510.62 - 7938.72 USD in 2022)
Internal expansion slots on 400 and 500 series: three 64 pin slots, and one 96 pin slot
The 300 series did have a first-party two slot expansion system available as an optional upgrade
The expansion slots were often used for things like SCSI, MIDI, hard disks, and more advanced math coprocessors.
Fun little thing I learned researching this topic: the power requirements of the CPU were so slight that to properly reset the system, an Archimedes needed to be powered off for at least 30 seconds. EMF from the fan could keep parts of the CPU powered.
Olivetti/Acorn wanted to break into the business computer market, so while working on the ARM2, Acorn had opened a research center near Xerox PARC in Palo Alto. Specifically, they wanted to get a much closer look at Xerox’s WIMP model. They weren’t unique here. Both Microsoft and Apple had investigated the Xerox system. Looking at this system, Wilson and Furber realized they needed to have more advanced graphics and sound, and this is reflected in the hardware listed above. The team in Palo Alto were working toward an OS they called ARX. ARX was a preemptive multitasking, multithreaded, multi-user, microkernel operating system with a windowing system. This system was apparently quite late.
Sophie Wilson commented:
ARX was late, very late. It became a black hole into which we poured effort.
Acorn eventually pivoted to produce a different operating system. The result was Arthur, whose named is derived by “ARM by Thursday”. This was done in just 12 months complete with application software, BBC BASIC, and an emulator for the BBC Micro. Arthur was a single tasking, single user, operating system with a GUI. The OS was stored in a ROM, and the Archimedes could therefore boot to it nearly instantly. The system was quite svelte, and would run well on a system with only 512K RAM and 512K ROM. This early OS borrowed heavily from the 8 bit predecessor. System calls were similar, much of the interaction was via a command line text interface, and graphics were accessed in a way similar to the BBC systems.
Arthur OS wasn’t intended to be the last word in the OS for the Archimedes. Development continued after the launch of the machines, and the result of that continued effort was RISC OS. This system improved upon its predecessor in nearly every way. It was a multitasking OS. Graphics and sound were far more easily handled. Text in the GUI was fully anti-aliased. There was a font manager with scalable fonts. The command line interface wasn’t needed for much of anything. This system was available as an upgrade ROM for all Archimedes systems, and it shipped standard on the 540.
Continued Troubles, ARM2aS, and ARM3
All the while, Olivetti hadn’t fully solved the financial troubles that plagued the company. The Archimedes was fast. It had a good OS in the form of RISC OS. It was expandable. It had multiple programming languages available for it. Yet, it was too expensive for most home users, and it was compatible with neither the hardware nor the software of the business computing world. The Archimedes line up found its home with programmers and hobbyists. This meant reliable sales figures, but no were close to the smashing success that Acorn and Olivetti really needed. As a result of the disappointing sales figures, Olivetti had begun looking to sell off ARM before the ARM3 chip was even brought to market.
With no takers, ARM development continued at Olivetti/Acorn. The ARM2aS moved to a 1 micron process on CMOS. This version of the ARM2 also allowed for the slowing of the clock, even pausing the chip entirely. This widened the power saving advantages of ARM. ARM3 made a large step forward as well. Beyond having on-die cache, it was faster in context switching, it could pass semaphores between tasks, and it was still far lower in power requirements than all competitors.
The Last BBC Micro
In may of 1989, Acorn released the BBC A3000. This was effectively an Acorn Archimedes in a case more reminiscent of the BBC Micro.
This machine used an ARM2 at 8 MHz, 1 MB of RAM, 512K ROM with RISC OS and BBC BASIC. This machine did have a floppy drive built in, and it had a single expansion slot. There was an external expansion connector to which existing expansion cards could be fitted. There was also an expansion chassis provided by Pres, along with a monitor stand that housed a 5.25” and 3.5” floppy drive.
The A3000 sold well to the UK educational sector like every other BBC branded Acorn machine. Its popularity in that space helped it get some representation in the mainstream retail market as well. Sales in both segments were helped by having a lower price (£649 in 1989 → £1598 in 2022) than other Archimedes models.
ARM’s First Outings with Apple and Nokia
In the late 1980s, Apple’s research group was working on the Newton. Their initial plans were to use the AT&T Hobbit CPU, and this wasn’t working well for them. They needed a very low power requirement CPU that performed decently well. Naturally, ARM entered that discussion.
With Olivetti seriously wanting to part with ARM, negotiations began between Olivetti, Apple, and VLSI around forming a new business out of the ARM team which at this point consisted of Jamie Urquhart, Harry Oldham, Dave Howard, John Biggs, Tudor Brown, Mike Muller, Al Thomas, Pete Harrod, Lee Smith, Andy Merritt, David Seal, and Harry Meekings. Then the question rose of who would lead this venture? The name that came up was Robin Saxby. He had experience with silicon at Motorola, he was an old friend of Herman Hauser’s, and he had some startup experience.
For its share of the company (43%) Apple invested £1,500,000 (a bit over £4 million in 2022 → 4.9 million USD). Apple also got a board seat. VLSI provided design tools and support. Acorn/Olivetti transferred the ARM IP and the ARM development team, while retaining 43% of the new company. VLSI was the first licensee of ARM’s IP from the new Advanced RISC Machines Ltd. The company’s headquarters was an old barn in Swaffam Bulbeck outside of Cambridge. The first conference table was won in a coin flip. Saxby tried to get some desks and chairs, but he lost the game of billiards.
Apple then released the Newton with its ARM CPU in August of 1993.
Also in 1993, Nokia asked Texas Instruments for a chipset for an upcoming mobile phone. Based upon the performance requirements and power requirements stated, TI advised the use of an ARM CPU. Nokia initially shot that down. The cost of the unit would be high. In response, ARM designed the Thumb (ARM7TDMI). This chip included the ability to decode new 16 bit ARM instructions. The advantage of this being that a hardware manufacturer could use cheaper 16 bit memory. TI was the first licensee. The first product was the Nokia 8110 released in 1996. Next came the 6110, 3210, and the 3110. This was the big break. There have been more than 170 licensees of the Thumb with over 10 billion units shipped since 1994.
While few could have known, ARM had just cemented itself as the mobile processor for at least the next 30 years.
Back at Acorn
Machines of the Archimedes line continued until 1995 with the last units being sold in 1996. The machines sold “okay,” but didn’t achieve BBC Micro levels sales. The successor to the Archimedes line was the Acorn RiscPC.
The Acorn RiscPC was launched on the 15th of April in 1994,
1994 – Risc PC 600, 30 MHz ARM610 CPU
1995 – Risc PC 700, 40 MHz ARM710 CPU
1996 – 200 MHz StrongARM CPU upgrade
1997 – J233 StrongARM 233 MHz Risc PC
The RiscPCs had dual processors. You’d have a primary CPU or host CPU as well as a guest CPU. The guest processors were normally 486 class x86 CPUs, though there were also 586 class CPUs available up to 133 MHz. Accompanying software allowed x86 PC software to be run on the RiscPC via the guest CPU (assuming 486/586). There were two 72 pin SIMM slots supporting up to 256 MB of RAM. Video was provided by the VIDC20 with up to 2MB of VRAM. Driving the machine was RISC OS starting with version 3.5 on the first model, and ending with RISC OS 3.71. In the image above, the base machine is on bottom. The top with the rounded front is an expansion chassis referred to as a slice. Each slice contained two podule expansion slots as well as a 5.25” and a 3.5” inch drive bay.
At this point, ARM was seemingly quite successful. The Apple Newton, the Acorn RiscPC, and the DEC StrongARM were all early wins for ARM.
Acorn, however, still wasn’t doing well. In 1996, Olivetti began selling off its interest in Acorn. From the New York Times on July 2nd of 1996:
Olivetti S.p.A. of Italy said yesterday that it had sold 14.7 percent of Acorn Computer Group P.L.C. to Lehman Brothers Inc. on Friday. Lehman did not disclose how much it paid, but at current market prices the sale would have brought about L33.5 million ($52 million) to Olivetti, which has been posting losses.
The purchase, representing 13.25 million of the British computer company's shares, reduced Olivetti's stake in Acorn to about 31.2 percent from 78.5 percent two years ago. Lehman said it intended to resell the shares to investors.
By 1998, Olivetti had fully let go of Acorn. Acorn closed their desktop and workstation computer divisions the same year. The 90s had many purveyors of RISC workstations: DEC’s Alpha, Sun’s SPARC, HP’s PA-RISC, MIPS, Motorola, IBM, and even Intel. Competing against these turned out to be too difficult a hill to climb. They tried their hand at set top boxes. These failed to catch much market.
Acorn then renamed itself to Element 14. This was part of a strategy change where they were focusing on digital signal processors (DSPs). At this point, the company was worth less than their shares of ARM, and the company was bought by Morgan Stanley on the 1st of June in 1999. The Acorn Group was delisted, and the shares of ARM were distributed among the new shareholders of Acorn at Morgan Stanley. Through a series of later sales and acquisitions, the owners of Acorn’s hardware IP ended up being Broadcom, and RISC OS is owned and developed (up to the present day) by RISCOS Ltd. Today, you can get an open source RISCOS version for the Raspberry Pi which features an ARM CPU.
ARM Goes Public
While Acorn was suffering a slow and painful death, ARM was soaring high. In 1998, they had an IPO on the London Stock Exchange with the name ARM Holdings plc and became a billion dollar company. By this point, ARM was the number three player in RISC, ahead of Intel and Motorola, but behind MIPS. The extreme rise in the value of ARM made the shares valuable. As a result, in 1999, Apple sold off quite a few its shares generating around $1.1 billion. While the Newton didn’t succeed, ARM did. Apple made out quite well, and the money came at a crucial time. That same year, ARM’s erstwhile manufacturing partner, VLSI, sold to Phillips for 1 billion dollars.
The start of the next millennium saw some changes in the market. Java had become a thing. SoCs were becoming a thing. With a commanding lead in the embedded market now firmly established, ARM was quick to release just the right product, the ARM9EJ-S. This produced less heat than its predecessors, had a higher clock, and shifted to the Harvard architecture. The Harvard architecture meant non-unified cache (the result of which is an instruction fetch doesn’t evict data, because they don’t share cache). This chip supported both 32 bit ARM and 16 bit Thumb instruction sets, sped up multiply operations, included an MMU, and supported the direct execution of Java byte code. This chip has shipped over 5 billion units since introduction.
In March of 2001, the Game boy Advance made its Japanese debut with an ARM7TDMI. In June, Nokia released the 9210 Communicator which used a 52 MHz ARM9 and ran Symbian 6. In October, Robin Saxby chose to step aside as CEO, and Warren East took his place. Saxby remained as Executive Chairman of the board. On the 23rd of October in 2001, the iPod was released with an ARM7TDMI at its heart.
A somewhat amazing thing happened in 2002. Intel got on board with ARM. They released their first XScale CPUs. Intel had bought DEC’s semiconductor division which included StrongARM, and XScale was the result. This implemented the ARMv5TE ISA at 0.18 micron and eventually at 0.13 micron. The final version of XScale was 45nm. These would range from 100 MHz to 1.2 GHz over time. The XScale’s most memorable uses were in the RIM Blackberry, the Palm Zire, Tungsten, and Treo, the Sharp Zaurus, the Compaq iPaq, the Creative Zen, and the Amazon Kindle. The most charming, however, was the Iyonix PC, which loosely cloned the RiscPC and ran RISCOS.
In April of 2002, ARM introduced what would be the last of its classic cores, the ARM11. This was ARMv6 with the Jazelle acceleration that allowed the execution of Java byte code, an eight stage pipeline with out-of-order execution, better branch prediction, and vectored interrupts.
By 2005, ARM was the CPU for 98% of all mobile phones. That year also saw the introduction of the ARM Cortex-A (as in applications), Cortex-M (as in microcontroller), and Cortex-R (as in real-time) series. Cortex-A was meant as the successor to the ARM9 and ARM11 cores. The Cortex-M series was designed to be low cost and low energy for use in microcontrollers. The Cortex-R series were intended for use in real-time and safety critical situations like medical devices or avionics.
The ARM Cortex-A8 was a hit. It was such a hit that Intel sold XScale to Marvell in June of 2006 for $600 million as the non-custom A8 was just that good. November of 2006 saw the acquisition of PortablePlayer (a chip supplier to Apple for the iPod) by NVIDIA, and PP’s ARM cores became the basis for the Tegra SoC family.
On the 7th of January in 2007, Steve Jobs revealed the iPhone. This used an ARM1176JZ-S CPU clocked at 412 MHz, PowerVR MBX Lite 3D GPU, and 128MB of RAM.
AT&T had recently bought Cingular, and the iPhone was a gift to them. Lines had formed outside their stores waiting for them to open. This was a new thing. Lines had formed outside of stores for other technology releases, but not quite on this scale and certainly not for phones. Within the first 30 hours, 270000 iPhones had sold. After 74 days, 1 million iPhones had sold. If we compare the first iPhone’s sales to competitors, it wasn’t the biggest thing. Nokia was king in the land of phones. Motorola was the next in line. Despite that, Apple had just entered the market and hit the top 10 in the first year. The iPhone 3G was released in 2008, and Apple sold 1 million of them in 3 days. It was faster, cheaper, had GPS, and it had slightly better battery life.
Apple already had some history of custom logic, but they were gearing up for a bigger project. In early 2008, Jim Keller left P. A. Semi to work on a custom ARM SoC project for Apple. To gain more expertise more quickly, in April of 2008, Apple purchased P. A. Semi for $278 million reuniting Keller with his prior team. Apple signed a deal with ARM that year that gave them the right make derivative core designs. Apple later and rather quietly, bought Intrinsity. Intrinsity had been another RISC logic company that had previously worked on PowerPC and MIPS. As a result of the talent they’d acquired, Apple announced the iPad with A4 chip on the 27th of January in 2010. The A4 was also also in the iPhone 4 which released on June 24th of the same year.
ARM’s Performance Roots Come Back
Apple pulled ahead with the first 64 bit ARM SoC. This was the A7, an ARMv8 compliant chip with thirty one 64 bit general purpose registers, thirty two 128 bit floating point registers, 64K data cache, 64K instruction cache, a 1MB L2 cache, a 4MB L3 cache, clocked at 1.3 GHz, and built on a 28nm process. The A7 SoC also contained a PowerVR G6430 GPU that shared the L3 cache. This thing was also quite wide (for the time) with 6 decoders. It had four FPUs. It had two branch units, and two load store units. The A7 was in the iPhone 5S. Apple claimed this was a desktop class chip. In an interview with Lex Fridman, Jim Keller said the same.
The Exynos 3 Single, first introduced in 2010, when compared to the Apple A4 is quite similar. It’s a 1 GHz Intrinsity Hummingbird core and a PowerVR SGX540 GPU. This a 45nm product. Samsung shipped it in the Galaxy Tab in September of 2010. Samsung then released the Exynos 4 Dual. This was a dual core 1.2 GHz ARM Cortex-A9. It was paired with an ARM Mali-400 MP4 GPU.
The Apple A7 and the Exynos 4 were a sign that ARM architecture was capable of being a serious computing powerhouse for serious computing tasks… just like those x86 guys. ARM was returning to its roots where it offered competitive compute power with less electrical power.
In 2011, ARM introduced its big.LITTLE heterogeneous computing architecture. This allows a device manufacturer to package simpler, power efficient, lower performance CPU cores alongside power hungry, high performance CPU cores. The advantage being that when idling or doing simple tasks, the user gets some serious energy savings. As memory is shared for all cores, tasks can be swapped between cores as needed. Additionally, as the less power hungry cores are also physically smaller, more cores can be packed onto a single die allowing for higher parallelism.
In 2013, Simon Segars became CEO after Warren East.
In 2014, the Riken Center for Computational Science in Kobe Japan began work on a new super computer. This computer uses the Fujitsu A64FX CPU core, which is an implementation of the ARMv8.2A architecture with a scalable vector extension, and 512 bit vector implementation. This is packaged with 32GB of HBM2 capable of 1TB per second. The CPU has 48 cores. The super computer at Riken uses 158976 of these CPUs resulting in 7630848 total ARM cores. The result of this at completion of the machine in June of 2020 was 416 petaFLOPS, which was quickly improved to 422 petaFLOPS, and then eventually to 2 exaFLOPS. This is Fugaku, a computer that bested the combined performance of the next 4 most performant supercomputers. Fugaku was the first ARM super computer.
On November 10th of 2020, Apple announced a transition to their M1 SoC. This uses ARMv8.5-A CPU cores in a big.LITTLE configuration. The M1 SoC and its variants surprised the market with performance rivaling the best that the x86 makers could muster. Similar to earlier designs, Apple went wide: 8 decoders, and a very large out-of-order design. There are many other innovations in the design, and after many years, ARM was back on the desktop with a high performance crown.
ARM Ownership Issues
On the 5th of September in 2016, SoftBank Group Corp completed its acquisition of ARM Holdings plc for 31 billion USD. This was a 40% premium over ARM’s share price. SoftBank agreed to keep the headquarters of the company in Cambridge, and also to continue to expand the UK workforce.
SoftBank invested a lot of money in ARM, and pushed to expand ARM’s presence in China. On the 4th of June in 2018, SoftBank announced that they’d agreed to sell a controlling interest in their Arm Technology (China) Co subsidiary forming a joint venture with the “Direct Investors” in which SoftBank retained 49% ownership. This joint venture represented the only legal way to sell or develop ARM IP in China. The SoftBank press release did not disclose who these “Direct Investors” were, and they didn’t disclose why the sale of 51% of ARM China was priced at a mere $775 million. One can only assume that this was a move by the CCP.
In 2020, ARM China chose in a 7-1 vote by the board to remove Allen Wu as the CEO of ARM China due to some corruption on Mr Wu’s part. Mr Wu disagreed. Allen Wu remained in power. In August of 2021, he announced the intention to take ARM China completely independent.
Following this announcement, Mr Wu hired security to enforce his claim, fired all employees who disagreed with his move, and then sued ARM China on the grounds that his firing was illegal. The lawsuit was especially humorous since Wu represented both sides of the case… he sued himself?
With all of this, ARM China rebranded as 安谋科技 (ARM Technology).
NVIDIA was an early ARM licensee, and had been using ARM technology in a variety of products for quite some time. On the 4th of September in 2020, SoftBank Group and NVIDIA announced the intent for NVIDIA to purchase ARM for $40 billion. In SoftBank’s press release, the hype was around AI technologies. This was planned to continue despite the dust up in China. The UK’s CMA and the USA’s FTC both had issues with this transaction.
The FTC Director Holly Vedova said:
The FTC is suing to block the largest semiconductor chip merger in history to prevent a chip conglomerate from stifling the innovation pipeline for next-generation technologies. Tomorrow's technologies depend on preserving today's competitive, cutting-edge chip markets. This proposed deal would distort ARM's incentives in chip markets and allow the combined firm to unfairly undermine NVIDIA's rivals.
On February 8th of 2022, SoftBank and NVIDIA announced the termination of the acquisition despite “good faith efforts by the parties,” and SoftBank said that ARM would IPO in 2023 instead.
Also the 8th of February of 2022, Rene Haas became CEO of ARM Ltd. Simon Segars resigned after 31 years and 3 months with the company.
As of this writing, over 230 billion ARM chips have been shipped.