Searching by SSE performance?

[schelles77]

We have recently discovered that our CV/ML application works much better on a Celeron N5095A than an N5105. When we benchmark both the only difference is in SSE performance - 1933M Matrices/s with the 5095A and 1533M Matrices/s with the 5105. That makes sense, because the ML uses the SSE instructions a lot, and a 20% difference is huge.

Is there any way to search the benchmark scores to find lower-end processors with high SSE values?

[David (PassMark)]

Not from our web site no. But in theory we could write a custom database query to do it.

But I think the issue is more complex that you suspect. The extended instructions / matrix test is comprised of 4 sub tests that use SSE, AVX, FMA and AVX512 groups of instructions.

So you might find a CPU with a really high score for the extended instruction test, but that high score might be from FMA/AVX instructions, but then you find your app doesn't support these newer instructions and you are stuck using the older SSE instructions when running your apps.

Also you might find a CPU with a good result where the good result is due to having a large number of cores, but your app might not be able to take advantage of such a large number of cores.

Also depends on what you mean by "lower end". Something like the Intel Core i3-10100TE @ 2.30GHz might be a good choice. (7.2M Matrices/Sec for around $100). So almost 4x the performance of the Celeron.