192xGT200 + 1068xBloomfield: 295 TFlops Supercomputer

French website PC INpact broke the news of an upcoming GPU-accelerated supercomputer, ordered by France's CEA for delivery in early 2009 from Bull. The cluster's performance confirms that GT200 will be rated at 1TFlop and that Nehalem/Bloomfield will clock up to at least 3GHz.

PC INpact claims that the overall machine will sport peak performance of 295 TFlops, with 103 TFlops coming from the CPU and 192TFlops coming from the GPUs. Le Monde further confirms the performance target and indicates delivery will take place in early 2009. This would make it one of the world's first (if not the first) large-scale GPU-accelerated supercomputer.

Sadly, PC INpact got the specifics wrong. There won't be a 8-core Nehalem in that timeframe, and neither could there be a 2TFlops single-chip GPU, especially given the qualification times in this market. Just like Conroe, Nehalem/Bloomfield will sport a 128-bit ADD and a 128-bit MUL per core; at 3GHz, that means 24 GFlops per core. Multiply that by the four cores per chip, then by 1068, and you get to ~103TFlops.

As for the GPUs, obviously the 1U Tesla module sports 4 GPUs, not 2. So assuming this supercomputer is indeed based on GT200 and the per-module config is similar, that gives us 1TFlop per chip/board. Alternatively, if the TDP was substantially higher than G80's (which seems unlikely given the level of binning possible in that market), it might be possible that the module only sports 3 GPUs; that would result in 1.33 TFlops per GPU. Once again, that is not the most likely scenario.

So, that gets us right back to where we were back in... May 2007, with Michael Hara's claim in an investor conference that their next-generation would be 'close to 1 TFlop'. Of course, we were all assuming that was referring to G92, and that obviously didn't turn out to be the case. We also assumed G92 would be the first chip supporting FP64; now, it has become clear however that's GT200 - which is also what makes the chip more attractive in such supercomputer deals.

In practice, this doesn't represent a lot of revenue for NVIDIA; it's very likely less than one million dollars for a company which is consistently delivering sales of more than one billion dollars per quarter lately. However, it does highlight Tesla's momentum in the GPGPU market. How soon, if ever, will that represent a substantial part of NVIDIA's profits? Nobody knows.