AVX Support in LAPACK functions

[Suneet]

Hi Intel, so I was pretty excited when I fired up my new Core i7 2600K and achieved 135GFLOPS inMKL linpackright off the bat.

Now I had read on MKL product pages that MKL LAPACK uses AVX, so I was expecting some performance boost for the parts of my science code that spend a long time solving eigensystems using ZSTEMR. But it was not to be! I ran my code and found only a small improvement.

So, I've now discovered that only a very few LAPACK and BLAS functions use AVX as of MKL 10.3.

My question is, can we expect more widespread support in the next version of MKL? Will AVX allow for a large (50-80%) speedup for solving eigensystems? When can we expect this stuff added to Intel MKL?

Thanks for any info!!

~Suneet

[Gennady_F_Intel]

Which the performance boost You are seeing on your system right now?

what is the typical size on your inputs?

--Gennady

[Suneet]

Hi Gennady, my usual problem size is 16384(^2) and double complex. So it is not very big and not very small, and we need to solve many of these.

Im solving for eigenvalues. With the Sandy Bridge Core i7 2600K at 4.8 GHz (amazing J) it is about 25% faster than 1x Xeon W5580 at 3.2 GHz. I think the increased performance can be attributed to clockspeed.

Usually the GFLOPS of the Intel Linpack benchmark is a good measure of how fast a system will solve my problems. However, this is not the case with Sandy Bridge because Linpack is using AVX double precision power and my MKL LAPACK functions are not AVX ready.

Is it even possible to speed up the eigenvalue solvers using AVX, something like 1.5x to 2.0x, the way Linpack has been sped up? Generally when can we expect for Intel MKL LAPACK to support AVX in most of the functions, and specifically some eigenvalue solvers? Perhaps the next few versions of MKL?

Thanks for any info on LAPACK support for AVX.

~Suneet