November 24, 2009

IBM Cuts Cell Loose

Michael Feldman

A few weeks ago, I speculated that IBM would probably not develop a follow-on Cell processor for the HPC market beyond its current PowerXCell 8i. A recent report by German publication Heise Online (English translation here) seems to bear this out. The article asserts that IBM’s plans for its next-generation Cell processor have been scrapped.

According to Heise, IBM VP of Deep Computing David Turek confirmed that there will be no successor to the PowerXCell 8i, the high performance Cell variant IBM developed for the Roadrunner supercomputer and its QS22 blades. Mercury Computer Systems, Fixstars, and Sony, also incorporated the processor into a few systems, but the technology was never widely deployed. At one time IBM had talked about expanding the Cell’s Synergistic Processing Elements (SPE) from 8 on the PowerXCell 8i up to 32 on the next-generation chip. That would have kept the processor on pace, performance-wise, with the latest and greatest GPUs from AMD and NVIDIA. Apparently though, this is not to be.

Ironically, the most energy efficient systems on the just-released Green500 list were all QS22, Cell-based machines. The three QPACE (QCD Parallel Computing on the Cell) systems in Germany took top honors with a rating of 723 megaflops/watt. That turned out to be nearly twice as efficient as IBM’s Blue Gene/P, the company’s other power-sipping architecture.

Turek implied that parts of the Cell technology will live on in future heterogeneous processors, or perhaps future stand-alone accelerators, developed by IBM. That, of course, leads one to wonder if he’s talking about variants of the Power CPU line, an upcoming Blue Gene architecture, or something entirely new. For the time being, Cell users can employ IBM’s OpenCL SDK to bridge their development to future systems.

If IBM pursues a heterogenous strategy, it will have plenty of company. NVIDIA, of course, is moving forward aggressively with its GPU computing roadmap, as exemplified by its upcoming Fermi products. Both Intel and AMD are pursuing heterogeneous architectures — Larrabee, Fusion, et al. — that incorporate GPUs or souped-up SIMD units on-chip. AMD also has its own graphics processor accelerator for HPC in FireStream, but has pursued a much more conservative strategy compared to that arch-rival NVIDIA.

In any case, the future of HPC is looking decidedly heterogeneous these days. Energy, space and cost issues all seem to be conspiring to force the CPU to play nice with more vector-like silicon. We should expect to see multiple flavors of CPU-GPU platforms, or their equivalent. It just won’t be the Cell.

[UPDATE: As implied in the original Heise article, IBM is only planning to ditch the Cell processor line for technical computing, not the original chip aimed at the Sony PlayStation gaming console. IBM says it is planning to keep churning out Cell silicon for that platform as long as needed.]