The GPU manufacturer held a first conference before the official opening of Computex, concentrating on GPU Computing.
NVIDIA announced the arrival of a server at SuperMicro equipped with an Intel platform and Tesla accelerators in a single 1U rack-mount server. NVIDIA’s own solutions are also in a single 1U but without the GPUs and must therefore be linked to a overarching server. The SuperMicro server then presents something new, though we don’t yet have the exact spec. All we can deduce from the 2 teraflops is that it will be equipped with 2 GT200b GPUs.
Jen-Hsun Huang, President & CEO, Nvidia.
In terms of the general consumer, NVIDIA talked up the growing number of apps that can use the GPU as a parallel coprocessor, at the same time saying that this is just the start as the arrival of DirectX 11 and Windows 7 should accelerate the trend. DirectX 11 means standardised use of GPUs (including current versions!) via Compute Shaders. Windows 7, as well as integrating DirectX 11, will include a transcoder that can use GPUs, via Compute Shaders, to accelerate video conversion.
Microsoft and Nvidia are highlighting native GPU support in Windows 7 for acceleration of video transcoding.
Thanks to CUDA, NVIDIA has a clear lead on the competition when it comes to GPU Computing because it has long been prepared for it on the software side. Although in the future all components should be supported, right now the software headstart will benefit the GeForces in terms of acceleration in certain video processing apps. The same goes for PhysX that benefits GeForces in terms of physics effects support.
NVIDIA is nevertheless grossly exagerating its vision and communication on what is an interesting subject is becoming something of a charicature. To convince the press that the GPU as coprocessor is the future of PCs for example, Jen-Hsun Huang, NVIDIA CEO, is saying that everyone is saying the same thing. Read, “you the press should write the same thing so as not to seem out of the loop”. Another example: when NVIDIA demonstrated a game with physics effects accelerated by the GPU, the sound volume was cut by half when these effects were disactivated, to accentuate the sensation!
What is most annoying however is the way that what is in fact a vector processing unit is presented as a core. This simplification is all the more annoying as GPUs and CPUs start to compete. If NVIDIA’s rhetoric is accepted then a quadcore CPU should be seen as composed of at least 16 cores if you take into account its vector units. Everyone would shout out how ridiculous this was and with good reason. NVIDIA has now put itself in a corner however as it has used this in its marketing and it would be difficult to backtrack. How could they say that the GT200 is in fact made up of 10 or even 30 cores and not 240?