2
$\begingroup$

I am looking into upgrading to a High Performance Computing work-station to do mostly computational fluid dynamics computations and some fundamental numerics. The CFD software that is available does not support GPU/CUDA. Aside from sufficient RAM and appropriate number of processors, the major bottleneck for my computations are the sockets of the processors.

There are two choices for processors: Intel and AMD. Intel interconnects their Xeon processors with QPI. AMD uses HyperTransport for their connection sockets.

Is there any reliable way I can compare the two just by using what is given to me in the data sheets? Has anybody here faced a similar decision? Advise is highly appreciated.

$\endgroup$
4
$\begingroup$

No there is no reliable way to predict the impact of interconnect characteristics on code performance solely from numbers on data sheets. One of the major confounding factors in the way of establishing the simple relationship you seek is the characteristics of your code about which we know nothing.

Yes, purchasers of HPC systems (large and small) face this decision all the time. The usual answer is simple: benchmark, in particular benchmark your code on the target platforms. If your purchasing power is too modest to bring a truck load of Intel and AMD kit and a truck load of Intel and AMD engineers flocking to your data centre, scare up some of the most similar hardware you can lay your hands on and do some of your own investigations.

I note that the article you pointed to in your comment on Wolfgang's answer is a report on benchmarking, not an argument from data sheet to code performance.

$\endgroup$
0
$\begingroup$

I think you are asking this question at too low a level. The processor interconnect is at such a fundamental level that you will have a hard time evaluating its effect. You also do not have much of a choice -- you can't just exchange this part of a system. Things that are by far more important (and that you can change by choosing different CPUs from the same vendor) are clock rate, memory bandwidth, cache size, how many cores, etc.

$\endgroup$
  • $\begingroup$ I disagree. Once you have reached a certain clock-rate, the interconnect does matter for shared memory and distributed memory computations. $\endgroup$ – seb Aug 13 '13 at 12:15
  • $\begingroup$ Forgot to add a reference to argue my point: here it is. $\endgroup$ – seb Aug 13 '13 at 12:21
  • $\begingroup$ I know it matters, but among all the variable that you can change, this one that you can't seems less important. You will be talking about single-digit differences in performance when clock rates, cache sizes, etc, will all combine to 20-30%. $\endgroup$ – Wolfgang Bangerth Aug 13 '13 at 14:52

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.