More Benchmarks Of AMD's Threadripper With LLVM Clang 6.0 SVN

[phoronix]

Phoronix: More Benchmarks Of AMD's Threadripper With LLVM Clang 6.0 SVN

With AMD a few days ago having landed an updated scheduler model for Zen CPUs within LLVM, I ran some fresh compiler benchmarks to see how the performance compares...

http://www.phoronix.com/scan.php?page=news_item&px=Clang-6-SVN-Tripper

[Kendji]

Threadripper is looking really good, though a bit out of my budget D:

[FireBurn]

Before and after the commit would have shown if there are any speed ups from it, there's so much churn all over LLVM & Clang it's not really comparable even though it does show the direction LLVM 6 is going in

[sdack]

How does one have to read the benchmark results?

I did look at the graphics, but when I'm looking at the comment boxes below them and where it mentions the compiler options did I only find two benchmarks that actually used either -march=native or -mtune=native. Were the other benchmarks indeed compiled with the default? What is the default here?

Also a lot of benchmarks apparently were only compiled with -O2 and not -O3. So what's going on there? Why leave the hand brake on?

The results themselves look terrible. Not sure if it's overall gotten worse or if it's just the same, but there is definitely no clear progress in terms of performance to be seen.

[Zan Lynx]

The ondemand scheduler is not good for benchmarking either.

[labyrinth153]

The goal of benchmarking is not always to confirm the maximum performance. Testing under usual conditions is excellent and sometimes more relevant.

[Zan Lynx]

The goal of benchmarking is not always to confirm the maximum performance. Testing under usual conditions is excellent and sometimes more relevant.

When I benchmark code on my laptop, the microseconds per call varies a LOT because of Intel power management. If I want reasonable, reproducible results I always have to lock it at performance and disable turbo.

AMD is even worse because their CPUs take longer than Intel's to shift power states under ACPI command. (The Ryzens have an in-CPU mode that is faster, but I don't think Linux ondemand drives it correctly)

[labyrinth153]

You are probably correct. I don't think I have any control over governors on my Mac.

[mlau]

The ondemand scheduler is not good for benchmarking either.

I found that on Zen, the governor doesn't mean much: the core frequency on the 1800x varies wildly between 1800 and 3900 mhz, depending on load,
regardless of the governor used. You can set the performance governor, but when the cores are idle, they're clocked to 1800mhz, which is even lower than
the lowest advertised frequency of 2200.