Intel Nehalem vs. Ice Lake Benchmarks - Including Clock + Power + Thermal Metrics

[ms178]

Is that a general statement or just for the Intel MKL-DNN (now DNNL) code ?

I wonder if the MKL_DEBUG_CPU_TYPE=5 environment variable (enable AVX on non-Intel CPUs) works for the DNN library ?

According to the following reddit post it should work everywhere where this library is used, but I cannot test it myself: https://www.reddit.com/r/matlab/comm...depath_on_amd/

I probably mentioned it before to you and it is unrelated to MKL-DNN but not enabling the fast path on AMD CPUs in Glibc is really in the same category: https://sourceware.org/ml/libc-alpha.../msg00155.html

With these sorts of artificial crippling out of the way, the performance on AMD only can get even better.

[bridgman]

If you're doing anything other than ML algorithms, apparently so.

Is that a general statement or just for the Intel MKL-DNN (now DNNL) code ?

I wonder if the MKL_DEBUG_CPU_TYPE=5 environment variable (enable AVX on non-Intel CPUs) works for the DNN library ?

[Michael]

So, is it your belief that all the relevant data files used by the benchmarks was read from disk cache (at least, to the extent it could've significantly affected performance)?

Anyway, thanks for the benchies.

BTW, on the outside chance you used sort to order your performance comparisons, you can specify -n to do a numerical comparison instead of lexical. Otherwise, that was a nice chart.

Aside from potentially the code compilation workloads, it's unlikely the HDD skewed the results by any substantive manner that would yield like a real measurable impact.

All of the graphs are generated by PTS/pts_Graph.

[coder]

Good thing I wasn't running storage benchmarks.

So, is it your belief that all the relevant data files used by the benchmarks was read from disk cache (at least, to the extent it could've significantly affected performance)?

Anyway, thanks for the benchies.

BTW, on the outside chance you used sort to order your performance comparisons, you can specify -n to do a numerical comparison instead of lexical. Otherwise, that was a nice chart.

[coder]

Too bad there isn't any Ice Lake desktop at least not yet.

True. Comet Lake is the next desktop line (also, high-end mobile) and will remain @ 14 nm. I'm not clear on the relationship of the uArchs, but we know that Comet Lake won't have AVX-512.

Anyway, Ice Lake will be coming to servers & featuring PCIe 4.0, in early 2020.

[stormcrow]

Haha, indeed. And it was pure scientific curiosity on my part. I guess people would and should spend their money with AMD on the desktop anyway at the moment.

If you're doing anything other than ML algorithms, apparently so. At least if you're going high end with Threadripper. With gaming rigs and general desktop I doubt anyone would notice if Facebook loaded 0.3 seconds faster or if you're system would render $GAME faster ... except that it's locked at 60 FPS anyway thanks to the engine being optimized for consoles instead of PCs like most popular games.

[ms178]

Too bad there isn't any Ice Lake desktop at least not yet.

Haha, indeed. And it was pure scientific curiosity on my part. I guess people would and should spend their money with AMD on the desktop anyway at the moment.

[Michael]

Tests already start wrong when you have one device using high-speed storage while the other uses a regular hard drive.

Good thing I wasn't running storage benchmarks.

[Michael]

A TDP-constrained environment plus modern software which uses AVX should be the worst case for the older architecture. I'd think that on desktop, let's say between Westmere @ 4.2 Ghz vs. Ice Lake @ 4.2 GHz (with all mitigations disabled) the differences shouldn't be as high (albeit with a similar trend). But at least in regards to gaming I wouldn't expect anything too drastic.

Too bad there isn't any Ice Lake desktop at least not yet.

[ms178]

A TDP-constrained environment plus modern software which uses AVX should be the worst case for the older architecture. I'd think that on desktop, let's say between Westmere @ 4.2 Ghz vs. Ice Lake @ 4.2 GHz (with all mitigations disabled) the differences shouldn't be as high (albeit with a similar trend). But at least in regards to gaming I wouldn't expect anything too drastic.