The applications so far, including the GPU version of protein prediction engine, were implemented using the ATI SDK 1.4 which uses Brook+ and have been tested on a 4770. Right now they *should* work on any card that supports gather and scatter streams which I believe are 3xxx and above. I say *should* because I found that ATI's implementation of Brook+ was not completely robust on Linux and I had to write around some "undocumented" features. Because ATI has moved to OpenCL, I intend to convert the routines to OpenCL which should be straightforward and should allow cross-compatibility with the NVIDIA cards.

For the most part, accuracy is not a problem with single precision (with the exception of one or two key variables that keep large sums and even there it coiuld be done in SP ). Actually, much of the work involved was converting the code to use 4-vector single precision math to exploit built-in optimizations.

For our own applications, the newer 40 nm cards (ATI 4770 and above and the next-gen NVIDIA cards) are where we are aiming at because of the low power consumption relative to TFLOPs. When running a cluster of machines 24/7, and not a single gaming rig, power consumption and heat dissipation are the limiting factors to the number of TFLOPs available.

Hong

Fix your link please cristipurdel , or which is the meaning of it, preview always before posting

Edit and link removed, tks

Thanks for the update.
Good news indeed.
And GPU version will be calculated on WGC platform or in house?
Wish you good luck wit project.h

Thanks for responce.
Wish you all the best with projest.
Please don`t forget on WCG since you will be not using our platform
and write few word about your progress.
We really look forward to it.

And is there any possible cooperation with HPF2 folloving Human Microbiome Project ?

@cristipurdel

The won't be a GPU calculations at WCG.They are doing it on their own computers