I have an S1070 box, connected via PCI-16 to a dual-quad Intel "nehalem"
cpu (2.93 GHz). Using the current beta MPI/GPU code and the benchmark
in amber11/benchmarks/jac/bench.jac.cuda (which is not the original jac
benchmark, but what Ross calls the "dhfr production nve" benchmark:
1 gpu 12.8 ns/day
2 gpu 19.5 ns/day (76% efficient)
4 gpu 27.4 ns/day (54% efficient)
For reference, using the cpus on this machine gives:
2 cores 3.2 ns/day
4 cores 6.2 ns/day
8 cores 10.5 ns/day
16 cores 11.9 ns/day
As usual, please remember that the GPU results are mixed precision, whereas
the CPU results are full double.
I think these are roughly comparable with what Ross and Scott have been
seeing, but they will know better than I do.
I'm hoping that my new box, with 8 C2050 cards, will be delivered soon.
...dac
_______________________________________________
AMBER-Developers mailing list
AMBER-Developers.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber-developers
Received on Tue Aug 24 2010 - 13:00:03 PDT
