When we wrote about DARPA’s UHPC press release earlier this week the notable thing was how little information it contained beyond the names of the lead institutions for the five teams. NVIDIA did its own release, spilling the beans about its team (Cray, ORNL, and others). Now we know a little more about the Intel team, thanks to teammate SGI.
Actually, that’s pretty much all the additional info we have. The SGI release says this
SGI…today announced it is partnering with the Defense Advanced Research Projects Agency (DARPA), Intel and other industry leaders to develop a future computing architecture that will overcome the limitations of the current evolutionary approach. DARPA’s Ubiquitous High Performance Computing (UHPC) program will develop an innovative, revolutionary generation of computing systems that will offer extreme performance and energy efficiency to deliver exascale supercomputer systems that meet the escalating demands of advanced computing.
“SGI is pleased to partner with DARPA and Intel on this groundbreaking program that promises to have far reaching implications for the supercomputing industry,” said Dr. Eng Lim Goh, senior vice president and chief technology officer at SGI. “To achieve the objectives of the DARPA UHPC program, we need to rethink how we are designing and building supercomputers. We look forward to a collaborative research environment that will result in major advances that redefine our industry.”
So we know that Intel has SGI and “other industry leaders” on the team. That locks up Cray and SGI — wonder if anyone partnered with IBM or HP? The SGI/Intel partnership isn’t exactly suprising, considering that SGI has been a stalwart partner of Intel’s for years, even through the period of AMD performance dominance when SGI was flogging its Itanium-based shared memory solution. The partnership could produce an interesting architecture, with Intel’s manycore development platform and SGI’s hardware shared memory glue. For example. And Intel’s giant software development team could certainly make a dent in all of the programmability requirements.