Since there are no export restrictions on single-precision performance, it wouldn't matter how large a behemoth is.bruce wrote:Drifting off-topic, I suppose . . . but what classes as a behemoth?
I was going for a random term to describe the typical class of hardware that falls under broad export regulations, which in the U.S. tends to be limited to the HPC sector, but which can vary by the exporting country.In the U.S. we specifically gauge theoretical DP-FLOPS against certain limits based on what kind of hardware is executing those FP operations and on the export recipient. The regulations have been revised repeatedly upward. There was a time when a Mac Pro (and every lesser-known 2P+ workstation of similar performance) required a license to export to most countries. That now doesn't happen. You now need 833 gflops on vector hardware and 2.5 tflops on scalar hardware, all in DP terms, to trigger broad export restrictions. That's like 32 Tesla C1060's!
Course, there are much stricter regulations for some countries where nuclear proliferation is a much bigger threat. Good luck sending even one i7 chip to N. Korea!
You can run Linpack on your folding machine whenever you want.7im wrote:Fah is almost exclusively 32 bit, yes, but the hardware it runs on is not. Wouldn't it be possible to extrapolate?
Running linear algebra on a GPU is another story - you'll perhaps need to code for it. Unfortunately the FAH client stats do not give us enough info to determine the aggregate double-precision floating performance of the active computers.... The other major problem is that all our folding clients are loosely connected across the Internet. If we were to run one giant instance of Linpack, the score would end up terribly low due to the interconnect latency.