News - AMD needs more than Shanghai Opterons to win server battle

[HEXUS]

AMD creates buzz around Shanghai at SC08, but the company's window of opportunity is small.

Read more.

[aidanjt]

Opteron has HPC performance issues?

[directhex]

Opteron has HPC performance issues?

Yes.

[aidanjt]

Yes.

Is that one of those "Jesus, I don't know what to do with all this compute power" performance issues?

[directhex]

Is that one of those "Jesus, I don't know what to do with all this compute power" performance issues?

Float performance is crap compared to Xeon. It's a problem Barcelona was meant to fix, and didn't

[aidanjt]

Float performance is crap compared to Xeon. It's a problem Barcelona was meant to fix, and didn't

Does this fp math disadvantage mean with the FPU, or SSE?..

[directhex]

Does this fp math disadvantage mean with the FPU, or SSE?..

With the FPU. With Santa Rosa, float performance is about half (!) woodcrest, per cycle per core. Well, on paper anyway (obviously in real life it's not that big a gap). Barcelona increased the theoretical float performance from 2 FLOPs/cycle to 3 - but due to design cock-ups, about 20% of that total performance is lost. Xeon has consistently had 4 FLOPs/cycle (as long as they're the right kind of FLOPs, anyway) since Woodcrest landed, with only the same memory bandwidth problems Xeon's always had causing issues. On purely RAM-bound codes, Opteron DOES tend to do better.

[aidanjt]

With the FPU. With Santa Rosa, float performance is about half (!) woodcrest, per cycle per core. Well, on paper anyway (obviously in real life it's not that big a gap). Barcelona increased the theoretical float performance from 2 FLOPs/cycle to 3 - but due to design cock-ups, about 20% of that total performance is lost. Xeon has consistently had 4 FLOPs/cycle (as long as they're the right kind of FLOPs, anyway) since Woodcrest landed, with only the same memory bandwidth problems Xeon's always had causing issues. On purely RAM-bound codes, Opteron DOES tend to do better.

Seriously?.. Odd, I thought AMD were already up to 4ops/cycle in native mode. Hrmm...

Anyway, you can't really wince at the aforementioned Cray XT5, they just keep scaling.