Strange FLOP counts

[eelsen]

Here are my results for different number of cores:

Matrix size 1024x1024:

No. cores    GFlops

1            0.062 (varies a bit from run to run)

2            0.093

3            0.102

4            0.103

Matrix size 512x512:

No. cores    GFlops

1            0.42

2            0.84

3            1.06

4            1.37

I used OpenMP to parallelize this matrix multiplication code. My experience is that

a 2X speed improvement in going from 1 to 4 cores is typical  for large problems.

Would SSE do much better?

[snapback]342838[/snapback]

If you’re just multiplying matrices - why not use the MKL? I think it gets something north of 50 GFlops for a single precision matrix multiply with 4 cores.

[nasacort]

If you’re just multiplying matrices - why not use the MKL?  I think it gets something north of 50 GFlops for a single precision matrix multiply with 4 cores.

[snapback]342854[/snapback]

I didn’t know that. I’ll have to try that some time.

Of course, I was also stupid in using the textbook formula Sum_k A(i,k) * B(k,j).

What I should do is to construct the transpose matrix D(i,j) = A(j,i) and then

compute Sum_k D(k,i) * B(k,j). This way, my CPU results look more respectable:

No. of cores GFlops

   1              0.72

   2              1.48

   3              2.22

   4              2.92

Thus, GFlops scales linearly with no. of cores and, furthermore, is more or less independent

of matrix size.