HPC

Intel Larrabee letdown leaves HPC to Nvidia's Fermi

Not so discrete graphics co-processors

By Timothy Prickett Morgan • Get more from this author

Posted in HPC, 8th December 2009 20:13 GMT

Free whitepaper – Operationalizing Information Security:

Comment Intel has never been particularly precise about what its "Larrabee" graphics chips were, so it is difficult to be sure how disappointed we should all be. And considering the company's track record outside of the x86 and x64 chip racket - its failed networking business and Itanium are but two examples of its woes - it's hard to say what to expect from Intel when, and if, it finally markets discrete graphics chips that can also be used as number-crunching co-processors for servers and workstations.

At the SC09 supercomputing trade show in November, when Justin Rattner, Intel's chief technology officer, gave the keynote presentation and his finale was to demonstrate a Larrabee co-processor being overclocked so it could hit 1 teraflops of single-precision floating point performance, there was no hint that Larrabee was on the rocks - the demonstration was clearly meant to show quite the opposite. That made the abrupt change last week that turned Larrabee from a graphics chip that was expected some time next year to a "software development platform for internal and external use," as Intel's statement put it, all the more jarring.

Advanced Micro Devices just keeps getting cut break after break, doesn't it? First Intel settles all outstanding antitrust lawsuits with AMD for a cool $1.25bn, allowing AMD to pay down some debts and clean up its books. And now Intel is delaying its entry into the discrete graphics market at the same time that IBM is admitting that it will no longer develop new PowerPC Cell co-processors for use in its own server lines.

The delayed entry of Intel's Larrabee and the dead-ending of IBM's Cell (at least on blade servers) gives AMD's Firestream GPUs a better chance against Nvidia's technically impressive Fermi family of Tesla 20 GPUs. The Fermi chips will be available as graphics cards in the first quarter of next year and will be ready as co-processors and complete server appliances from Nvidia in the second quarter. And they will likely get dominant market share, too, particularly among supercomputer customers who want to have error correction on the GPUs - a feature that AMD's Firestream GPUs currently lack.

But still, a market always wants at least two alternatives, even if it rarely wants more than three, and that means AMD still has time to get ECC onto its Firestream GPUs and compete head-to-head with Nvidia's Fermi GPUs.

Next page: Larrabeen or Neverbee?