by
Anonymous Coward
on 2007年03月24日 22時52分
(#1131600)
GPUの数値が良いわけ↓がこれだそうで
Just counting FLOPS can be deceptive. The GPU does a lot of FLOPS, but is less efficient with them.
For example, the GPU doesn't allow a scatter (random access write) so we are forced to do some calculations twice:
the force on i & j is related to the force on j & i, but saving that data requires a scatter, so we just recalc.
That's the faster way to go on the GPU, so it's the right thing to do there. It also brings the FLOPS *way* up,
but it means it *requires* 2x the FLOPS just to do the same calculation.
スペックの情報が欲しいなぁ、、、 (スコア:2, 興味深い)
せめて CPU のクロックくらいは欲しいなぁ、、、と、、、
とりあえず、単ノード当たりの GFLOPS 値を比較してみた。
uxi
Re:スペックの情報が欲しいなぁ、、、 (スコア:1, 参考になる)
Just counting FLOPS can be deceptive. The GPU does a lot of FLOPS, but is less efficient with them.
For example, the GPU doesn't allow a scatter (random access write) so we are forced to do some calculations twice:
the force on i & j is related to the force on j & i, but saving that data requires a scatter, so we just recalc.
That's the faster way to go on the GPU, so it's the right thing to do there. It also brings the FLOPS *way* up,
but it means it *requires* 2x the FLOPS just to do the same calculation.