During the technical session on designing the future of supercomputing, presenters from TACC (Texas Advanced Computing Center) went into detail on creating the Stampede supercomputer, world’s first supercomputer powered by the Xeon Phi coprocessors. The system targets compute power of almost 10 PFLOPS, with the majority of compute power coming from Xeon Phi.
The system originally targeted 10 PFLOPS, but it seems they might miss the mark by a few dozen/hundred TFLOPS. According to information given, the Stampede deploys 2 PFLOPS of compute power through Sandy Bridge-EP based Xeon CPUs and no less than 7 PFLOPS using Xeon Phi "coprocessors".
Even though TACC did not disclose how many thousands of Dell servers are being deloyed, we know that the supercomputer has 272TB of DDR3 memory and 14PB of total storage. The Xeon processor is E5-2680 (Sandy Bridge-EP), while TACC is using special versions of pre-production Xeon Phi processors in order to make the launch.
When it comes to the actual Xeon Phi card, Intel still hasn’t disclosed the performance and features. However, we can reveal that the first Xeon Phi cards will feature 62 cores, 8GB of GDDR5 memory (connected to a 512-bit bus), using PCIe Gen2 (FirePro and Tesla utilize PCIe Gen3). The TDP is set at 300W,
Furthermore, Center decided to pack 128 Tesla K20 (GK110-based) for visualization purposes. While this deal represents a major loss to NVIDIA, it is telling that Xeon Phi cards need Tesla in order to visualize data. Will this be the case with all future Xeon Phi supercomputer wins – time will tell, but NV and probably AMD (FirePro is finally available in server variants) could end up riding Intel’s co-tail and achieve some revenue (albeit in percentile figures).