I’ve been watching today’s news out of ISC2011, as my research group is firmly planted in the field, and it’s a hell of a shakeup industry-wise.
The most important fact is that the age of “GPUs are the future!” has come to its (as we’ve been saying, well and truly due) end, based on two specific claims:
The big news is that the #1 supercomputer has turned over to a half million processor Fujitsu-built SPARC64 box, at Riken in Japan, leapfrogging the GPU-based #1 from November’s Top500. SPARC based systems haven’t been doing well of late, and Oracle’s acquisition of Sun made it seem like things were going to get worse, so this is truly a machine out of left field. It also appears to be doing unusually well in term’s of its Rpeak/Rmax ratio (which is to say, how close to it’s nebulous theoretical performance number it can get on a benchmark), which has been a major problem for GPU based machines.
The other impressive news out of ISC11 is Intel showing a new Larabee generation called Knights Corner, with 50 smallish x86-64 (w/ AVX) cores in a GPU-like package. Unlike the 32-core Knights Ferry parts that have been circulating in research environments for the last few years, Intel is indicating these will be widely commercially available as part of the Xeon line by 2012, and support double precision floats in hardware. Far more important, their initial test users have been showing effective ports of existing parallel programs in hours or days, rather than the full rewrites in CUDA (proprietary) or OpenCL (nasty) process required for GPUs. Incidentally, Intel is also claiming to have a native OpenCL tool chain for the new chips, and coupled with The Portland Group’s cuda-x86 tool, should be able to run CUDA code as well.
To put it another way, Nvidia is pretty much done in the supercomputing market.
This actually has some ramifications for the research group: on the upside, it has the very nice side effect of validating the need for our MOG tool to make GPUs an attractive target, which will hopefully attract some collaboration and funding for the project as various entities who bought into the GPGPU thing begin grasping at straws. On the downside, the fact the US only has one machine in the top 5, which means most of the US government supercomputing funding (around $2.5B a year – about a one fighter jet’s worth of spending) is going to be redirected into a mad rush to get things on the top of the list — basically, a big buy from Cray instead of research funding.
It should be a fun year for HPC.
