STRONG ARM TACTICS

Dick Pountain/14 September 2010 10:14/Idealog 194

In the interests of transparency I begin this column with full disclosure: in 1992 I accepted a bottle of 21-year Springbank single-malt whisky from Robin Saxby (now Sir Robin) the CEO of Arm Holdings Ltd. It was an innocent enough gift. I'd just written a feature for Byte about the ARM610 processor which Apple was considering building into the Newton, and I ended saying that if this tiny UK chip maker could clinch a deal with the US giant I'd celebrate with a dram of Scotland's finest, which to my genuine surprise turned up with a ribbon around it.

ARM is a very British success story, in that 99% of the population have never heard of it, it doesn't actually manufacture anything, yet 15 *billion* of its chips are fitted into mobile telephones, cash machines, medical instruments, toys and so on, throughout the world. And the firm may present more of a challenge to Intel's market dominance than AMD.

The original Acorn RISC Machine (ARM) launched in 1986 to power Acorn's Archimedes PC, using a minimalist architecture designed by Steve Furber, who is now ICL Professor of Computing at Manchester. Pretty soon the ARM team came to two imporant conclusions - no British firm could hope to compete with Intel in manufacturing chips, but their tiny simplified processor core could be shrunk further, more easily than even the smallest of US RISC designs. Accordingly they decided to sell only intellectual property - processor designs - to other chip makers, in return for licence fees and royalties.

So small was the ARM CPU core that even by the 1990s it was possible to put more than one onto a single silicon die, and ARM diversified the design into various dedicated cores for memory management, disk control, signal processing, communications and so on, so that by the late '90s it was possible to build the electronics for a whole device onto a single die that consumed far less power than competing chips. They sold the rights to these core designs to huge manufacturers like Samsung, Qualcomm and even to Intel, who added extra logic of their own and made the actual chips. Chips containing ARM cores, and thus running ARM binary code, now dominate the mobile phone marketplace. The Snapdragon processor fitted in many Android touchscreen smartphones? Qualcomm, with ARM core. The iPhone? Powered by Samsung chip with an ARM core. And many, many more ARM cores live inside embedded controllers that manage everything from cars to cash machines.

However dominance in the mobile and embedded markets alone is never likely to threaten Intel, which has the huge desktop, laptop and servers markets almost to itself, yielding only a small slice to AMD since Sun quit the server scene. ARM's next moves could be about to change that, because the superb shrinkability of that core puts it in harmony with two of today's leading concerns in the server market: energy saving and virtualisation.

As the volume of data stored in the world's search engine indexes and databases continues to mushroom, datacentre power consumption becomes of critical concern: scare stories continually surface about the internet consuming most of the world's electric supply by the year 20xx. Now the ARM may appear to suffer a tremendous disadvantage in the server market because it remains a 32-bit architecture and thus limited to addressing no more than 4Gbytes of memory directly. What's more, revamping it completely into a full 64-bit design might well spoil its excellent scalability and power-consumption advantages. Processor designers for years have solved such dilemmas by applying virtual memory, a memory management unit that translates each memory access into a larger address space and fools the CPU into only seeing one 4Gbyte chunk at a time (Intel has done it since the first Pentium). That has disadvantages both in operating system complexity and power consumption.

However the server world nowadays is less and less interested in operating systems per se and more and more interested in hypervisors. A typical web server is going to be virtualised, with each processor core running several virtual machines, each machine serving one request with fairly modest memory requirements. It's only applications like video editing that actually demand terabytes of contiguous memory nowadays. Not too much surprise then when in August this year ARM's architecture program manager David Brash revealed that the ARMv7-A (aka "Eagle" and "Cortex") features hardware extensions to address a terabyte or more of memory in two stages, and also virtualisation extensions that create a new privilege level for suitably designed hypervisors running over the ARM core. In September ARM revealed that its next-but-one Cortex product will feature 16 cores on the same die, running at 2.5GHz, and will be aimed at the cloud server market.

Assume that there's also a communication controller core in the pipeline with equivalent capabilities to scale I/O bandwidth, and you'd have a solution to those problems of parallel processing I've been writing about here recently, which might use an order of magnitude less power than current architectures and give Intel a real battle at last. If that comes to pass I'll celebrate with a glass of 1949 La Tache and a 300gm tin of beluga caviar (worth a try, worked last time...)

[BIO: "Dick Pountain started out on Burgundy, but soon hit the harder stuff"]