AMD Ryzen 9 5950X + GCC 11 Compiler Benchmarks At Varying Optimization Levels

[phoronix]

Phoronix: AMD Ryzen 9 5950X + GCC 11 Compiler Benchmarks At Varying Optimization Levels

Following last month's initial benchmarks of the AMD "znver3" support that landed in the GCC 11 compiler was a request by a premium supporter to see the AMD Zen 3 benchmarks at more compiler optimizations. Well, here are those numbers for those wanting to pursue aggressive compiler optimizations on a shiny AMD Ryzen 9 5950X.

http://www.phoronix.com/vr.php?view=29836

[sp82]

This numbers are just unbelievable.

[ext73]

Yeah I already give it and it gives our users much more efficiently By the way, you can now use the ZEN 3 instructions on ZEN 2 - it works better than zen2 on zen2

https://www.dropbox.com/s/vmlhvhz7u0ija0f/Screenshot_20210107_224325.png?dl=0

[mb_q]

This numbers are just unbelievable.

Yup, time to switch to GFLOPs...

[carewolf]

No geometric means graph?

Anyway cool mix of options. Hopefully the compiler writers can learn something from the at times odd results.

[sn99]

I would love to see a clear linux optimized for amd vs normal clear linux on amd

[AmericanLocomotive]

Absolutely ridiculous how much performance we're leaving on the table in the name of compatibility. Especially with applications that have huge gains from modern instructions. AMD and Intel go through back-breaking amounts of effort, combing through every last piece of their hardware trying to squeeze out 1% extra performance where they can.

...while we're leaving ~13% extra performance on average laying on the table just by targeting x86-64 instead of a modern architecture... That's nearly a processor generation of performance. Many applications showed a greater than 25% increase in performance - that's a greater performance increase than any generation-to-generation CPU performance increase in the past ~6-7 years or longer.

if x86 wants any hope at competing with ARM they're going to have to cut grandma with her Core 2 Duo loose.

[jrch2k8]

Absolutely ridiculous how much performance we're leaving on the table in the name of compatibility. Especially with applications that have huge gains from modern instructions. AMD and Intel go through back-breaking amounts of effort, combing through every last piece of their hardware trying to squeeze out 1% extra performance where they can.

...while we're leaving ~13% extra performance on average laying on the table just by targeting x86-64 instead of a modern architecture... That's nearly a processor generation of performance. Many applications showed a greater than 25% increase in performance - that's a greater performance increase than any generation-to-generation CPU performance increase in the past ~6-7 years or longer.

if x86 wants any hope at competing with ARM they're going to have to cut grandma with her Core 2 Duo loose.

1.) yeap

2.) x86_64 is everything, the problem is generic

3.) the same problem applies to ARM btw but is less noticeable because ARM is not very mass friendly and everyone build their own uber SoC optimized version(starting by the fact not 1 SoC boot the same way).

Also replace generic is not very simple without some GlibC/libstdc++/etc. dark magic because the main issue is Intel and AMD extension support wildly vary between architectures and vendors and the performance cost of those extensions are not equal even between the same vendor.

For this we would need so sort of automatic FMV (function muti versioning) and FatElfs + the dark magic

[FireBurn]

Wow, you've finally listened and produced some really interesting benchmark results of useful compile options

All that was missing was -march=native

[900k]

Absolutely ridiculous how much performance we're leaving on the table in the name of compatibility. Especially with applications that have huge gains from modern instructions. AMD and Intel go through back-breaking amounts of effort, combing through every last piece of their hardware trying to squeeze out 1% extra performance where they can.

...while we're leaving ~13% extra performance on average laying on the table just by targeting x86-64 instead of a modern architecture... That's nearly a processor generation of performance. Many applications showed a greater than 25% increase in performance - that's a greater performance increase than any generation-to-generation CPU performance increase in the past ~6-7 years or longer.

if x86 wants any hope at competing with ARM they're going to have to cut grandma with her Core 2 Duo loose.

I dont think that its as simple as dropping older cpus / enabling new instruction sets, ive found that even just setting the mtune option to "intel" improves performance even on older cpus like the core 2 chips in at least some cases.

I would be curious to see what the impact of different mtune options would be across multiple architectures,