Furthermore, "The CG kernel figures out the approximation to the smallest eigenvalue of a large sparse matrix"-- that is not really correct. CG is a common iterative method that tickles some of the right buttons in terms of sparse matrix structure, non-coalesed memory access, etc. While it can be used to estimate the eigenvalues of a matrix, that really is not why it is included in the NAS benchmark suite.
As with every single MIC paper I have seen, this one completely avoided making any direct comparisons to GPUs. Likely because the Phi co-processors, despite being targeted as a competitor to GPUs, are not really competitive with them yet. I'm not criticizing the paper, which really was about efficient schedulers, but the article for trying to make an article about something it is not.
I like nextplatform but this article was not up to the typical quality standards.
You're going to see fewer and fewer comparisons to GPUs, since Intel has finally figured out that users don't want to cram all their data into a PCIe lane. The current generation of KNL is not really a coprocessor -- it's self-hosted, so instead of a crapload of dual-socket Xeon processors with MIC cards hanging off them, you have a crapload of single-socket KNL processors, on which users can run x86 code directly. nVIDIA is still superior, in my eyes, but they're going to need to get nvlink figured out and deployed, because Intel has basically started deploying accelerator cards with no CPU layer needed -- and no recompile, either.
2 comments
[ 0.23 ms ] story [ 10.7 ms ] threadFurthermore, "The CG kernel figures out the approximation to the smallest eigenvalue of a large sparse matrix"-- that is not really correct. CG is a common iterative method that tickles some of the right buttons in terms of sparse matrix structure, non-coalesed memory access, etc. While it can be used to estimate the eigenvalues of a matrix, that really is not why it is included in the NAS benchmark suite.
As with every single MIC paper I have seen, this one completely avoided making any direct comparisons to GPUs. Likely because the Phi co-processors, despite being targeted as a competitor to GPUs, are not really competitive with them yet. I'm not criticizing the paper, which really was about efficient schedulers, but the article for trying to make an article about something it is not.
I like nextplatform but this article was not up to the typical quality standards.