Intel may claim that its Xeon Phi GPU compute accelerator cards are powering the world’s most efficient HPC data center, but the reality is that the efficiency is not coming from the Knights Corner architecture at all.
During this year’s IDF event in San Francisco, California, the company has displayed a few systems powered by the new architecture and the power consumption numbers are not rosy at all. We’ve detailed here that the efficiency in data centers is mainly dependent on intelligent cooling systems as well as on an efficient computing architecture. Intel’s new x86 based Xeon Phi is just as efficient as AMD’s old VLIW GPU computer cards and the company itself is stating this in its presentations as we reported here. We consider this as a successful first attempt from Intel to produce a capable and efficient GPU compute accelerator, but this is hardly a fitted competitor for Nvidia’s K20 or AMD’s GCN.
In an interesting development, hardware experts at Semiaccurate.com have managed to snatch some pictures of Intel’s Xeon Phi demonstrations at IDF 2012 and the power consumption numbers revealed makes us wonder about the touted “efficiency.” Intel’s Xeon Phi, in its 56-core version, consumes an amazing 258 watts when idling, SemiAccurate reports. There are two cards in the displayed system and one is reported as consuming “just” 190 watts when sitting at idle. That’s a huge amount considering that AMD’s VLIW accelerators idle somewhere below 40 watts while the GCN generation cards have an idle below 10 watts. There are not many details available about the cards right now, but even the 58-watt difference between the two adapters, that are supposed to be absolutely the same, is making us wonder about manufacturing and yield problems with the new chip.
Intel Xeon Phi IDF 2012 Presentation Image credits to SemiAccurate |
0 comments:
Post a Comment