Performance of MKL FFTs on Westemere and Sandybridge

[subramanian__vedapra]

Why is the performance of MKL FFTs on Sandybridge node is 3-5 times faster than Westmere node?

I could think of these 2 reasons

1. Sandybridge has more CPU cores than Westemere

2. May be MKL is smart enough to use the AVX support available on Sandybridge.

Can anyone eplain why the performance of MKL differs between Sandybridge and Westmere?

[SergeyKostrov]

>>...Can anyone eplain why the performance of MKL differs between Sandybridge and Westmere? I'd like to request some additional information. Could you provide exact models of CPUs for both systems? For example, if you're on a Windows platform msinfo32.exe utility displays that information. In all the rest cases you could use a CPUid utility. Thanks in advance.

[Bernard]

>>>Why is the performance of MKL FFTs on Sandybridge node is 3-5 times faster than Westmere node>>>

As you stated in your post it is very likely that improvement is due to newer architecture.The main improvemnt can be wider 256-bit units so you have twice FP throughput.If the uops do not contain interdependencies Port0 and Port1 can both at the same time issue one AVX mul and one AVX add operations.This can be combined at the same time with the two load/store opeartions performaed by the Port2 and Port3.Nehalem/Westmere was able to issue only one load and one store operations on its Ports.

[SergeyKostrov]

>>...2. May be MKL is smart enough to use the AVX support available on Sandybridge... There are lots of details related to performance improvements of MKL in Release Notes: software.intel.com/en-us/articles/intel-mkl-110-release-notes but, unfortunately, I didn't see any notes related to applications of AVX for FFT.

[Evgueni_P_Intel]

Vedaprakash,

Please provide details regarding FFT lengths, type of signal (complex/real), computing environment (cluster/single computer), and the processors that you are comparing (e.g. "Xeon E7-8867L", "Core i7-3820QM", etc.)

Thanks,

Evgueni.