Tianhe-1A installation - 7168 NVIDIA Tesla M2050s

Let me put it this way, if parallel and scalable algorithms for a particular problem are in widespread use, then it is reasonably likely that the problem would do decently well on the gpu. With the imminent introduction of rw caches, the gap is going to fall even more.

Out of curiosity, it would be interesting to see bzip/7zip/whatever ported to gpu. It might be able to do reasonably well on large datasets, (time spent in data transfer aside).

 

Interestingly enough, but I think the driving force for GPUs in HPC wont be Nvidia or even AMDs Fusion, but Intel's Knights technology.

1. Larrabee is as good as it gets for GPU HPC workload. x86 compatibility, easy programing, porting and much higher efficiency than either Nvidia or AMD solutions. Maybe it wont beat flops records, but its positives outweighs negatives IMO.

2. Larrabee will be fused with Xeons for the max performance.

3. Intel's reputation and influence, along with IMHO superior HPC products, will make it obvious choice for supercomputers, thus Nvidia and AMD wont play a major role in HPC area. It seems AMD isnt investing that much in supercomputing, but it will hurt Nvidia a lot, since its losing not only HPC, but GPU market as well.

Another heavyweight player - IBM with Power CPUs and Cell combo, but programming for it is even more cumbersome than for Nvidia, therefore IMO the main player with "gpu's" in HPC area will be Intel, and in not so distant future.

 

Let's wait for it to get to market before we announce winners.

 

Your arguments do have some weight, granted. But not even Intel considers Knights a GPU, which I have explicitly mentioned in my post above. Nor seems IBM (or Sony) consider Cell to be one (RSX in PS3).