RealTime IT News

Intel's Larrabee Hits 1TFlop of Computing Speed

Larrabee, Intel's first big push into the graphics processor unit (GPU) market, didn't make much of an impression at the Intel Developer Forum (IDF) last September. However, the company made up for that at the recent SC09 supercomputing show.

The conference focuses on all things related to high performance computing, something GPU vendors have increasingly wanted a part of. ATI and nVidia were both on hand at SC09 to show off their products tuned for HPC, as was Intel. Only instead of just discussing the Nehalem generation of CPUs, Intel CTO Justin Rattner also showed significant progress in Larrabee's development. (See YouTube video clip here).

Intel has been coy about the exact specs of Larrabee. It has not said how many cores it will have, the clock speed or power draw. What is known is that it's a many-core design that uses older Pentium cores, updated with 64-bit extensions and SIMD extensions plus a 1,024-bit high-speed ring bus to link all of the cores.

The IDF showing was, to be charitable, underwhelming, with an old 3D graphics game sitting idle on stage and barely any action taking place on screen. But during his keynote, Rattner showed a much better demonstration. He held up the Larrabee card, which is about as large as modern day cards, and then ran an HPC demo.

Using a SGEMM single-precision math performance test and half the cores on the card, Larrabee racked up 417 gigaflops (GFLOPS) of performance. With all the cores turned on, it hit 825 GFLOPS, which would indicate 1:1 scaling that Intel has said it is aiming to provide. With a little overclocking, Rattner got it to hit the 1 teraflop (TFLOP) barrier.

By way of comparison, the Intel quad-core QX9775 Core 2 Quad, the highest-end processor of the pre-Nehalem generation, taps out at 51.2 GFLOPs. The Core i7-975, the top desktop of the Nehalem generation, hits 55.36 in turbo mode and 42.56 in regular mode.

Ready for primetime?

Good, says Jon Peddie, president of Jon Peddie Research, but not good enough. "You've got ATI out with a card [the 5800] that can do five teraFLOPs now. For Intel to come out with a card that does one teraFLOP next year isn't going to cut it in the high-end space," he told

He qualified that by stating the part does well, but has to compete against the very established players in ATI and nVidia. "I think Intel wants to show the progress they've been making, and the part can deliver performance, and one teraflop is a good level of performance. If they price it according to other one teraflop parts from ATI and nVidia, then they have a contender," he said.

The question is how much of a contender. "Intel will take market share from nVidia and ATI. How much remains to be seen. Intel will get market share just for showing up, but going into the performance market they will have work to do. They are going into nVidia's and ATI's camp where there is tremendous brand loyalty," Peddie added.

That said, he feels Intel is doing things right, and certainly not fudging the numbers. "I don't think Intel would play any games. It's a big publicly held company they can't afford to do smoke and mirrors," he said.

"They have a program plan and as far as I know they are executing close to plan. They've done the seriously hard work in terms of getting the developer community onboard. The hard work they had to do was all the software tools," he added.