ARM Holdings of Cambridge, England, claim that that the cumulative total of the processors manufactured under their license has reached 25 billion. ThatÂ´s three-and-a half Turing machines for every man, woman and child on the planet.
Turing machines? Yes. These are not special-purpose controller chips, but general-purpose programmable computer processors.
They are in principle capable of running any digital algorithm you can come up with, the definition of a Turing machine, and they can therefore emulate each other. Some algorithms never stop, and others wonÂ´t stop before the universe runs out of heat, but thatÂ´s a petty detail. From the great height of a mathematical logician like Alan Turing of KingÂ´s College, Bletchley Park and (to BritainÂ´s shame) Manchester police station, all implementations of his machine are the same.
Of course if you add the desiderata useful and reasonable time, processors are every much not equal, and faster is better. We have advanced from the huge, crude, snail-like Colossus and ENIAC that respectively merely broke OKWÂ´s Lorenz ciphers – a step beyond Enigma – and designed the H-bomb, to todayÂ´s inch-square marvels that can precisely render the glinting helmet on the snarling head of an orc, decapitated by the sword of your heroic avatar, as it rolls gratifyingly across a virtual dungeon floor. Orc credit
ARMÂ´s processors (and those of its competitors) are embedded in mobile phones, MP3 players, cars, TVs, routers, set-top boxes and a raft of other gadgets. Even its first and simplest model, the ARM 2 marketed in 1986, ran at 8 MIPS, two-thirds of the contemporary Intel 386, workhorse of early PCs. The latest perform at over 5,000 MIPS per core, against the 30,000 or so MIPS per core of IntelÂ´s.
From one perspective, itÂ´s overkill to use the equivalent of 160 ENIACs to drive your car GPS or washing-machine. However, itÂ´s now clearly cheaper to put an overspecified general-purpose computation engine in a controller than to design and debug an optimised one from scratch.
YouÂ´d be very hard put to get these embedded cores to run generic programs, though perhaps some enterprising Indian is building a Beowulf computer out of discarded mobile phones. The processors are typically not even distinct physical devices, just segments of larger chips, like Texas InstrumentsÂ´ OMAP series, nestled with other components like graphics, sound processors, and USB controllers.
These Turing machines have become computational mitochondria: the bacteria that opted, more than a billion years ago, to live permanently within eukaryotic cells. The deal is that the mitochondria provide the host cell with energy (ATP) in exchange for food (glucose) and shelter (the cellÂ´s wall and cytoplasm). The mitochondia in the cells of a particular animal are all the same, though they power many different types of cell, from muscle cells to phagocytes to neurones. They are SFIK pretty similar between species.
We like to think as mitochondria as the humble slaves of their host organisms, but thatÂ´s just our native eukaryotism speaking. You could equally well see the hosts as elaborate gentlepersonsÂ´ clubs for mitochondrial comfort. The commensal contract is a fair exchange. Do you own your cat or does your cat own you?
Mitochondrion photo source
Similarly, ARMÂ´s microscopic generic computation devices power a great range of much larger chips and devices. They are beginning to multiply on chips: a mobile phone controller used to have one CPU, one new design has five – still a long way short of the hundreds of mitochondria in a typical cell. Unlike mitochondria, you can still just see CPUs with the naked eye: 1/2 to 2 millimetres on a side, the thickness of a hand-drawn line. Where the analogy breaks down is the pace of evolution. Mitochondria avoid the hazards of sexual reproduction by pure maternal inheritance. Accordingly their design changes slowly over the aeons, loosely coupled to the evolution of the hosts. Mitochondrial DNA still looks very like that of bacteria, with compact ring-shaped chromosomes. Computer processors are changing under our eyes, driven by MooreÂ´s Law.
ItÂ´s a striking oddity that the market for computer processors, the heart of the competitive global electronics industry, is highly monopolistic. Intel has manufactured the great majority of all the mainline, CISC microprocessors ever made. It has usually had competitors, but no more than one or two at a time: Texas Instruments, Motorola, IBM, Cyrix, AMD. In the low-power market, ARM has had things more or less to itself [update: this is overstated – see Dave JacobowitzÂ´s comment infra], though as mobile and desktop devices converge, Intel and ARM are becoming competitors of each other.
These monopolies have been pretty benevolent. The network economies of standardisation to their direct and indirect customers are huge. ThatÂ´s why we stick with Windows, and Windows sticks with Intel. Learning to run a different operating system is a considerable investment cost to an end-user. Learning to interface a graphics card or operating system with a different processor architecture and instruction set is a very much bigger one. So thereÂ´s one huge barrier to entry.
However, itÂ´s not an absolute one. Processors cannot possibly be marketed as black boxes.The intermediate customers are highly knowledgeable, and need to be given an enormous amount of information about the processors to be able to work with them. Intel and ARM know that if they bring to market dud devices, or attempt to gouge on prices, or slow down on the innovation treadmill, their potential competitors would tunnel through the externalities barrier and become real ones – or, in ARMÂ´s case, to attempt a takeover, as itÂ´s small compared to many of its customers. Like the Red Queen they cannot stop running.
ARM`s business model is quite different to IntelÂ´s. Intel is an integrated design-and-build manufacturer. It sells boxes, and likes to see its sticker on computers that use them. ARM cares nothing for name recognition and has never sold a box since the demise of the BBC Acorn before 2000, a nice little educational computer in its day, with a prescient name. ItÂ´s a pure design shop. The designs are licensed, in immense detail, to device manufacturers all over the world, at around 6 US cents a copy. An example from a recent trade press release:
Taiwan-based United Microelectronics Corp. (UMC), one of [sic] worldâ€™s largest semiconductor foundries, is now the exclusive supplier of Kindle Fire ARM processors. UMC is going to supply to Amazon Texas Instruments OMAP4430 through the 45-nano process. The OMAP4430 represents a dual-core 1GHz processor produced on ARM architecture ….
The Kindle processor intimately involves ARM in Cambridge with three customers: Amazon in Seattle as end customer and designer of the complete device; Texas Instruments in Austin as designer of the complete controller to AmazonÂ´s requirements; and UMC in Taiwan as the merchant chip foundry. The latter two are certainly licensees; I donÂ´t know if Amazon also is. One has to marvel at the ability of the Internet to abolish distance in this very high-level technical collaboration. ThereÂ´s no question that the final product works, and reliably so. Has the coffee-shop physical proximity of Silicon Valley become unnecessary for its successors?
IÂ´m rooting for ARM for one reason beyond chauvinism: power. Its <$10 chips use milliwatts compared to IntelÂ´s traditional tens of watts. Power is increasingly a bigger constraint than performance. In MIPS, ARMÂ´s processors roughly track IntelÂ´s of a couple of years back. Still, its next generation will run 64-bit Windows, in direct competition with Intel for laptops.
Also, and this is more important for global warming, servers. Running a Google search on your ARM-powered iPad uses little power in your hand, but adds to the 258 MW of GoogleÂ´s ever-growing server farms. These are basically racks of hard drives and hot Intel motherboards. Departing from its usual policy of inoffensiveness, ARM has co-funded a startup in Texas to make server cards using its processors. We should wish them luck. Anyway, real competition – even in a duopoly – will be even better for customers: and very likely the planet.
BabbageÂ´s 1871 Analytical Engine of 1871 was the worldÂ´s first mechanical programmable computer. This is just a part that got built. It never really worked, so its delivered MIPS were 0.