The current uptake in high performance computing means mostly good things, but it also comes with a few built-in challenges. The paradox of this particular progress is this: when you scale hardware, you oftentimes scale power consumption, right along with it. That’s where Intel’s Shesha Krishnapura has some good news to share, in this podcast speaking with The Register’s Tim Phillips. Says Krishnapura, “In the past, that power relationship has existed. But with Intel’s core microarchitecture platform, the power holds constant while performance climbs.”
Intel is working to improve the performance-per-watt characteristics of HPC systems. The effort is important, as Xeon-based servers dominate the Top 500 supercomputers list and the clusters used by businesses for their most demanding jobs.
Fist of all, Intel’s throughput-per-rack measurement helps illustrate the point when Intel 45nm-based quad-core processors run at similar power levels as dual-core processors, while offering twice the number of processing cores per server. Add Intel’s switch to higher density memory like 4GB memory modules instead of 2GB modules — the 4GB run at similar power envelope — and it’s clear where Intel is holding a fairly stable power envelope and still seen what Krishnapura calls, “a substantial performance increase, year after year.”
(more…)