nVidia is turning on the heat on its GPGPU line-up: after launching an add-in card, 1U server part and the Tesla Personal Supercomputer, the company decided to launch a program for their partners which now include the whole cluster.
As a case study, nVidia is touting the case of Bank National du Paris-Paribas [BNP Paribas], where the bank decided to throw old setup out – 500 CPU sockets sucked 25kW of power on a continuous basis. This was replaced with a smaller cluster consisted out of several CPUs feeding two Tesla S1070 systems – for a total system consumption of 2kW. We wonder how, given the fact that each of Tesla S1070 systems features four GT200-based Tesla cards. AMD Opteron-based server with 16GB of memory will eat 0.4kW on its own, meaning that eight 55nm GPUs and 16GB of GDDR3 memory consume the remaining 1.6 kW.
The odd part about this case study is that BNP Paribas didn’t use a Tesla GPU Preconfigured Cluster, since nVidia specifies that a "Tesla GPU Preconfigured Cluster" is consisted out of "x86 CPU servers coupled with Tesla S1070 1U GPU systems. Configurations start at 16 TFLOPS of performance, delivered by four Tesla S1070 servers with 16 GPGPUs and 64GB of GDDR3 memory. All systems include host servers, Infiniband switches, cabling and are fully customizable to suit individual needs." It definitely smells like another case of overly-enthusiastic PR department or as Jon Stewart would say, a clusterf#%k http://www.thedailyshow.com/ . BNP Paribas deployed HALF of the "Tesla GPU Preconfigured Cluster", daamit.
In any case, BNPs "Mini-Me GPU Cluster" resulted with 190x decrease in power consumption, yet alone to discuss radical reduction in cooling demands for the GPU cluster. nVidia claims that there are 15-20 million engineers that are struggling for computing time on supercomputers, claiming that the newly launched Tesla Preconfigured Cluster is the answer to their needs. Riding the CUDA train could pay off handsomely for companies in the Tesla Preconfigured Cluster eco-system. Seeing old friends from Appro and Colfax mentioned in combination with Dassault spells one thing on the wall: Formula One teams and automotive industry.
Given that next year’s season will feature a pretty stingy budget cap, don’t be surprised to see announcements between F1 teams and nVidia partners.