'Cell', the massively parallel processing chip currently being designed by Sony and IBM, will scale from single-chip systems through to entire server rooms packed with thousands of them, Sony's executive deputy president Ken Kutaragi told attendees of the company's Transformation 60 conference yesterday.
That's always been the goal, of course, since Cell was first announced back in March 2001. But yesterday Kutaragi put some numbers onto the chart.
A four-core chip home server system will be able to deliver one billion floating-point operations per second, apparently. Move up to a 32-core chip - in, say, a blade server module - and you'd get 32 gigaflops of processing power, while a 64-core slab of silicon inside a rack-mount unit doing graphics work would churn out two teraflops, according to Kutaragi's presentation foils.
Ultimately, Kutaragi suggested, we'll see 16 teraflop supercomputer 'cabinets' and one petaflop (a million billion flops, in other words) server rooms - the latter delivering enough raw power for true AI systems, he said.
Kutaragi likened a single Cell chip to IBM's 32-node RS/6000-based chess supercomputer Deep Blue. The exponential scaling rate suggests Cell really doesn't come into its own until you use lots of them together.
That's certainly the design philosophy: "With built-in broadband connectivity, microprocessors that currently exist as individual islands will be more closely linked, making a network of systems act more as one, unified 'supersystem'. Just as biological cells in the body unite to form complete physical systems, Cell-based electronic products of all types will form the building blocks of larger systems," was how Kutaragi described the Cell concept back in 2001.
Since it takes 2200 PowerPC 970 chips - aka the G5 - to yield just over ten teraflops - much the same as you get from 2000 Athlon 64s - getting similar performance out of just 64 Cell cores is impressive, if Sony and co. can deliver.
Right now, they're just a little way past half way through the five-year Cell research project, so they have a few more years yet to demonstrate the device in action. We'd expect the successors to today's chips to have got a little closer to those kinds of figures by 2006, but not that close.
Others are not so far behind. ClearSpeed's recently announced CS301 chip, for example, can deliver 25 gigaflops peak, the company claims. The CS301 is a 64-way parallel processing co-processor designed to work alongside an x86 or other general purpose CPU. The downside is that it's expected to cost over $1000 per chip. We suspect Sony and IBM are aiming for something a little more mass-market. ®