topic Performance on MKL and IPP in Intel® oneAPI Math Kernel Library

Performance on MKL and IPP

missing__zlw — Wed, 08 Sep 2010 21:27:13 GMT

I am sorry if this question has been asked and answered. Search didn't turn out valid answers.

I notice both MKL and IPP provide similar functions such as FFT, eigen computation. Is MKL a wrapper on IPP? Does MKL provide better/worse accuracy? Will I get better performance using IPP?
Ifmy typical FFT is a 400x200 complext matrix.

It seems MKL has info on processors MKLGetVersion.Processor. Does it optimize for different processor? Does IPP optimize the same way?

My bottom question is which one I should use? MKL or IPP?

Thanks.

Performance on MKL and IPP

Shane_S_Intel — Wed, 08 Sep 2010 22:29:55 GMT

There's a KB article on the subject - http://software.intel.com/en-us/articles/mkl-ipp-choosing-an-fft/

MKL and IPP are distinct and different libraries, although in the case of FFTs and VML they do share optimized kernels at the lowest levels. There are likely cases where MKL will give better performance and others where IPP will lead. Generally MKL focuses on performing well for mid to larger sized problems common in HPC; IPP's focus is smaller sizes common in the multimedia, data processing, and communications areas.

Both MKL and IPP provide optimizations fordifferent IA processors.

As for your last question, much depends on how you intend to use the libraries longer term. You should be able to download free 30-day evaluation copies of both and see which one best suits your needs.

-Shane

Performance on MKL and IPP

missing__zlw — Thu, 09 Sep 2010 04:24:17 GMT

Thanks for the article. It is very helpful.

So for a 400x200 complex 2D array, is it considered as mid-large size, or small size?