NVIDIA AND MICROSOFT have started pitching the former's Tesla as the überkit for high performance computing by giving developers a hand on the Windows HPC Server 2008 platform.
According to the new bedfellows, their plan is to encourage developers in adopting Windows HPC Server 2008 as their choice of development platform for parallel computing tasks using Nvidia's GT200-based (eventually Fermi too) Tesla systems, while Nvidia provides the developer community with the necessary software in Nexus. Nexus allows developers to work on massively parallel applications in a familiar Visual Studio-based development suite.
The partnership is still being fleshed out, but if we follow Jen-Hsun's lead on Day 1 of Nvidia's Nvision 2.0, er, GPU Technology Conference, you can see what it's building up to.
Nvidia has retained a solid reputation in the workstation and server business due to Quadro and the first Tesla "supercomputer" chips, however traditionally this kind of presence is accessible only to a handful of companies with very, very deep pockets.
Nvidia promises to address the main concerns with supercomputing installations - power consumption and performance - as scientific compute complexes and render farms sucking more power than small cities. This power and cooling problem can be resolved by plugging in Tesla instead of CPU clusters, it claims.
Nvidia spokesperson Kerry Tescher told us, "Tesla fits extremely well in enterprises' energy-saving attitude. GPU computing delivers 10x the performance in the same power budget."
She added, "A desktop system with Tesla GPUs delivers the same performance as clusters with 16 to 32 CPU servers. Similarly, a small Tesla GPU cluster can outperform clusters that are almost 10-20 times bigger and thus significantly save energy."
According to Nvidia slideware, 32 Tesla S1070 chips will draw 27 times less power than their CPU server farm counterpart, in an oil and gas seismic processing scenario. Even a 2.7 times improvement would be a breakthrough, so if the numbers are anywhere near reality it seems Nvidia might be onto a cash-cow here.
Ah, but where does Fermi fit in all this, you ask? Well, although Fermi has just been announced, the Tesla S1070 and C1060 GT200-based GPUs are still driving Nvidia's HPC strategy. Tesla will make use of Fermi down the line, but right now all the numbers handed out are based upon the current GT200 architecture.
Fermi's increased programmability and GPU-as-CPU DNA should make things a bit easier for developers, especially the C++ crowd. While that hasn't happened, Microsoft Research and Nvidia are busy deploying a "large Tesla GPU computing cluster", although Nvidia seems to be only holding Microsoft's hand in this, as it provides no numbers on computational power or the profile of the installation.
Cray, Dell, HP and Lenovo are lining up for some orders, according to sources. µ
Give Glass a pass if you're not an enterprise type
Epyc performance potential, but Rome wan't build in a day
Patch? Patchy more like