topic Re: How to get/compute achieved GFlops Performance in Intel® oneAPI Math Kernel Library

How to get/compute achieved GFlops Performance

matthiasd — Thu, 27 Nov 2008 21:13:01 GMT

Hello!

I am currently using the Intel MKL for research activities.

For rough performance estimation I coded some benchmarks. But I am not sure how to compute the actual reached GFlops Performance. Due to the fact that I don't know how much actual floting point operations the blas function really executes.

(on the Intel Website I found benchmark graph with dimension on x-axis and GFlops on the y-axis.)

here is an example from my results

Intel Q6600, 2.4 GHz, 2 GB RAM
cblas_sgemm (C <- alpha * A * B + beta * C)
matrix sizes of 3000x3000 (random values, full filled)
computing time: 3.350 seconds

I am also interessted in some hard facts about the Q6600 like

peak GFlops Performance
cycles for a float/double operation (i guess several float ops per cycle with SSE, Pipelining, ...)

Thanks for your anwsers. I would by really nice to get reliable data at first hand.

Re: How to get/compute achieved GFlops Performance

TimP — Thu, 27 Nov 2008 22:41:38 GMT

There's little question about the number of floating point operations in matrix multiply, given that MKL is a work-alike for public BLAS, but plenty of useful reading accessible with a few clicks in your browser.

Noting that Core 2 Quad is identical architecturally to predecessor versions of Core 2, except for number of cores, if you are serious about research, you should refer to the processor manuals http://www.intel.com/products/processor/manuals/

The primary ratings for cycles per operation are throughput (msximum retirement period) and latency (serial period).

Linpack Top 500 ratings refer to efficiency (floating point operations completed)/(theoretical peak). Beyond that, such calculations have little practical use. http://www.top500.org/ Are you starting a category of "hard, but nearly useless" facts?

Note that Core 2 Quad is labeled as Kentsfield on the Top 500 site. 5420 of them to achieve a score of 32800 Gflops (double precision).