Speed-ups

[UH-HPCTools]

What should be the speed-up we should get by using accelerators?
I tried Matrix multiplication code using accelerator. But only got 3-4 times speed up!

[Tiger]

The speed-up's seen when using an accelerator will vary depending on what you are doing. You will typically see a peak-performance quoted for an accelerator chip, The ClearSpeed CSX700, for example, has a peak performance of 96 GFlops for Double precision.

As an application example, when the host calls a DGEMM (matrix mulitplication) a single CSX700 chip can provide 77 GFlops. The ClearSpeed DGEMM can use the host processor in conjection with the CSX700 ("host-assist") which enables the CSX700 performance to be "added" to the host performance.

Also the CSX700 is available in a CATS box, which holds up to 12 accelerators. Each accelerator will be capable of added additional performance.

So a 3-4 times speed-up can be classed as either very good or quite average depending on your algorithm, the size of the problem, the capabilites of your host, the type of accelerator chip and the number of accelerator chips.