Related News

Pure GPU Based HPC Coming to an End?

iSGTW published an article discussing the future of NVIDIA's CUDA supercomputers in light of Intel's and AMD's next generation chips

Both Intel and AMD currently manufacture new processors that integrate graphics-related processing capabilities right on chip: Intel's "Sandy Bridge" and AMD's "Llano" processor - both to be shipped mid-2011. As the article states, it is not to be expected that further graphics capabilities are required. Since integrated chips come at a much lower price than full graphics boards, the author concludes that NVIDIA will no longer be able to maintain the low cost for CUDA it's currently offering. Read the full article here
Whilst the argument of cost is certainly a fair one with respect to low-end graphics needs, such as for notebooks, the issue looks slightly different for High Performance Computing, where in particular vector like processing capabilities find more and more uptake. In the modern Top 500 list, hardly any supercomputer will not incorporate larger amount of GPU nodes. True, both Intel and AMD are working on bringing vector capabilities onto the chip, such as Intel's AVX technology, but until these reach a scale comparable to modern day graphics cards, NVIDIA remains in a good position in this area.
We will certainly see the rise of more convergence, i.e. of tighter integration of GPU and CPU in the future though.