The Flop (Floating Point Operations per Second) Rate of MATLAB Code

Hello, I know Intel MKL / IPP libraries performance in simple operations (Multiplication, Summation, Matrix Multiplication, Vector Multiplication) gets something like 80-95% of the theoretical performance of the CPU (Measured in FLOPS).

http://software.intel.com/en-us/articles/parallelism-in-the-intel-math-kernel-library http://software.intel.com/en-us/intel-mkl

Yet, doing so using MATLAB I get much worse results.

I have this simple script:

numElements = 2 ^ 16;
numIter     = 100;

vecX = randn(numElements, 1, 'single');
vecY = randn(numElements, 1, 'single');

initTime = tic();
for ii = 1:numIter
    vecX .* vecY;
end
stopTime = toc(initTime);

gFlops = (numElements * numIter) / stopTime

Yet I get only 1.1 GFLOPS on my i7-860 Which should be closer to 2.8GHz (Frequency) * 4 (Cores) * 4 (Single Precisio Operations per Cycle as SSE Vector - 128 Bit) = 44.8 GFLOPS.

Yet I get something like 1.4 GFLOPS. Which is only 3% of the theoretical performance.

How can MATLAB be so inefficient?

[Amit]

BTW, MATLAB is only using 1 core, I'd believe. And for a benchmark, is there anything else running besides MATLAB?

[Royi Avital]

MATLAB use more than one core.

Answer 1

[Amit]

I don't think the way you're trying to calculate flops here is right. Even if one assumes that you can calculate Flops like this, you're missing out many overheads that matlab is doing. For example, try something like this:

numElements = 2 ^ 18;
numIter     = 100;
vecX = randn(numElements, 1, 'single');
vecY = randn(numElements, 1, 'single');
initTime = tic();
% for ii = 1:numIter
%     vecX .* vecY;
% end
vecX + vecY; % I used +, but you can switch to .* as well
stopTime = toc(initTime);
gFlops = (numElements * numIter) / stopTime

And see if you see any difference. I am pretty sure you will. Remember, for loop is slow.

[Royi Avital]

@Amit, I tried, actually the same results (Though you must remove the 'numIter' factor in your code). It seems MATLAB handles the loop pretty well.

It just generates slow code which I don't understand why since it uses MKL.

I think anyone suggests something, should try it and see his results compared to the theoretical peak of its CPU.

[Amit]

I didn't see that I have used numIter (my bad :P)

You can try Lightspeed toolbox http://research.microsoft.com/en-us/um/people/minka/software/lightspeed/

which has a module for flops and see if you get the difference.

[Salma Hassan]

can i use tic and toc only to compue the time elpase

matlab 2018b

[Walter Roberson]

tic and toc only provide elapsed time information, which is not the same as the amount of computation done, as elapsed time can include time that the operating system suspended MATLAB in order to work on something else.

Answer 2

[Walter Roberson]

single() is often slower than double()

Your arrays are not that big; I am not sure that it is kicking in calls to the libraries.

[Royi Avital]

I tried with doubles and larger vectors. Same result.

Feel free to try yourself and show results with better utilization of the CPU.

Such a pity this software isn't close to take a real advantage of the resources.

[Walter Roberson]

Try with timeit. Or if you have an older MATLAB that does not have that built-in, you can get timeit from the File Exchange.