本帖最后由 扫帚 于 2012-8-3 18:58 编辑
Albums: We have the exact production specifications, such as the final clock speeds, not yet published. It comes naturally to the particular program - the question is, limits the number of functional units or the memory interface? Limited by the latter, you can probably expect a speed gain of about 50 percent. Compute-intensive applications such as SGEMM likely, however, quite good scaling with the number of units.
EU: So you could suggest an average speed advantage between 50 and 70 percent? Albums: This or something more you could hope for ordinary programs.
你所说的80%在哪里呢?
|