Initial Benchmarks With Intel oneAPI Level Zero Performance

[phoronix]

Phoronix: Initial Benchmarks With Intel oneAPI Level Zero Performance

Last week Intel released an initial set of micro-benchmarks for their oneAPI Level Zero and with L0 support being plumbed into their open-source Intel Compute Runtime, this weekend I started toying around with some Level Zero benchmarks on a variety of Intel processors.

http://www.phoronix.com/vr.php?view=29045

[alpha_one_x86]

Hi, seam you have compare Intel oneAPI Level Zero with Intel oneAPI Level Zero, on multiple hardware. Right?
If right, can you compare same software native vs Intel oneAPI Level Zero?

[coder]

Phoronix: Initial Benchmarks With Intel oneAPI Level Zero Performance

Thanks for these.

I have a few questions about the units:

• Peak Integer Compute should be rated in GOPS, not GFLOPS
• Device-To-Host Bandwidth & Host-To-Device Bandwidth - are the usec scores simply the time taken for the test execution? Are these not basically redundant with the GB/sec score (except also useless, since we don't know the amount of data being transferred)? If they're not redundant, then certainly more information is needed to divine any insight from them.
• Host-To-Device-To-Host Image Copy - I'd love to know why this is consistently so much faster than Host-To-Device Bandwidth. Smells like either a measurement error, or an opportunity to learn something interesting.

Thanks.

[Michael]

Thanks for these.

I have a few questions about the units:

• Peak Integer Compute should be rated in GOPS, not GFLOPS
• Device-To-Host Bandwidth & Host-To-Device Bandwidth - are the usec scores simply the time taken for the test execution? Are these not basically redundant with the GB/sec score (except also useless, since we don't know the amount of data being transferred)? If they're not redundant, then certainly more information is needed to divine any insight from them.
• Host-To-Device-To-Host Image Copy - I'd love to know why this is consistently so much faster than Host-To-Device Bandwidth. Smells like either a measurement error, or an opportunity to learn something interesting.

Thanks.

They are all the units and values reported ny the upstream level-zero-tests from Intel.

[coder]

They are all the units and values reported ny the upstream level-zero-tests from Intel.

Thanks for the answer, but is there someone at Intel that you could forward these questions to? Maybe you could feature their answers in the article, during the next round or iteration of oneAPI tests.

Thanks.