Replies to post #6257 on AMD strictly moderated (AMD)

[chipguy]

That's a good point (very few). But Barcelona may solidify AMD in the HPC market.

Too bad the HPC market is extremely varied and SPECfp_rate
is very non-representative of at least 80% of apps. HPC apps
can vary in their bandwidth usage from < 0.1 GB/s per GFLOP/s
to >1.0 GB/s per GFLOP/s. In many areas NGMA is way ahead of
K8 (bioinformatics and EDA for example) and more raw FLOPs
isn't going to help much.

[wbmw]

Re: Barcelona may solidify AMD in the HPC market.

I think AMD will try to target SPECfp_rate and certain HPC apps as if they were the Achilles Heel of the Xeon MP platform. But I think Intel will have plenty more workloads at their disposal to counter this tactic. At best, it will give AMD a means to market themselves better than they can today against Tulsa and Clovertown. But don't forget that Intel will be much better equipped to respond than they were in the days of Paxville and Irwindale.

Re: Another variable in the equation is that there are some code sequences that can be optimized through vectorization, and I am assuming that these SSE(2) instructions gain parformance proportionally with FP.

I wouldn't assume that.

[mmoy]

Core 2 has some significant SIMD improvements that have put it
ahead in multimedia performance. Even if you can process FP more
quickly, you still have to do load/store.

If you need DP, then you don't have a lot of options but if the
needed precision isn't that high, you can frequently get a lot
more work done with scaled integer.