The EE Times ran an article yesterday on a talk given by David Kirk, chief scientist at NVIDIA, at London’s Imperial College. In the talk he opined on the degree to which GPU technology stands to democratize supercomputing
“For $2,000 per teraflops, anyone can have a single-precision supercomputer in their PC,” he said. “Very soon we will have double-precision floating-point designs, and I predict that within less than two years you will see pentaflops [sic] clusters of double-precision floating-point designs for under $5 million.”
Pentaflops is, I assume, a corruption of “petaflops.” Kirk went on to describe his vision for HPC
At the University of Illinois, Kirk teaches a class for scientists and engineers on employing massively parallel processors for general computation. At present, around 100 universities worldwide provide courses based on the Illinois material. “We need to teach massively parallel computing to everybody in science, not just computer scientists and electrical engineers,” he said.
The piece goes on to talk about how Kirk sees CUDA developing and the impact it might have on the future of HPC.