Discussion Rudi_Float_Bench v0.02a

[igor_kavinski]

v0.02 with an additional AVX-512 specific binary (it will either crash or exit unexpectedly if the CPU is lacking the necessary ISA extensions): https://drive.google.com/file/d/12RuZsWdhNueu7th2HCuzblBA8CUGFu9u/view?usp=sharing

If it complains about missing vcruntime140 DLL something, install this: https://aka.ms/vs/17/release/vc_redist.x64.exe

Older version (not much point in testing this. Only putting it here for historical purposes): https://drive.google.com/file/d/1l7PU3W0u82iJovpbmJ9FTnhGMHJoVJGw/view?usp=sharing

Thanks, MS_AT and Hail for identifying the compiler issue.

And now for some scores!

 

[Markfw]

Well, what about your 64 core AMD Rome ??

 

[igor_kavinski]

Well, what about your 64 core AMD Rome ??

I'm not home yet

 

[MS_AT]

So what this benchmark is measuring, as for sure not what it claims, as those intrinsics are AVX512 specific, so no way it could run on anything but Xeon.

 

[igor_kavinski]

So what this benchmark is measuring, as for sure not what it claims, as those intrinsics are AVX512 specific, so no way it could run on anything but Xeon.

Visual Studio 2022 is apparently putting in an alternate AVX2 codepath. Seems they wisened up and did the right thing, for once.

The actual hot loop code:

 

[MS_AT]

Visual Studio 2022 is apparently putting in an alternate AVX2 codepath. Seems they wisened up and did the right thing, for once.

The actual hot loop code:

View attachment 119235

Have you verified avx512 ops are actually in the compiled binary? Because any sensible optimizing compiler should just discard them when higher optimization levels are used (O2, O3, might be even O1). And if you want to have a benchmark you should use at least O2 equivalent or higher.

Unfortunately godbolt is unusable on mobile so I cannot check.

 

[igor_kavinski]

Unfortunately godbolt is unusable on mobile so I cannot check.

Waiting for you to check coz I'm just a 44 year old kid

Please do enlighten us!

 

[igor_kavinski]

Well, what about your 64 core AMD Rome ??

Unfortunately, Ice Lake wins

Because:

The benchmark has some bug. It won't use more than 64 threads. I'll have to check the code tomorrow to see if there is a hard limit in there.

If Ice Lake is really using AVX-512, that might explain how 48 threads are able to beat 64 physical Zen 2 cores, though the Epyc puts up a fierce fight armed only with AVX2 and gets pretty close. Considering that the Ice Lake server cost $7000 and my Epyc cost not more than $1500, I would say the Epyc wins fair and square

However, if the compiler has "cheated" and replaced the AVX-512 code with AVX2 code, the only possible explanation for the Epyc's loss is higher frequency on the Ice Lake CPU (3.95 GHz vs. 2.91 GHz on the Epyc). But even then, with almost a GHz deficit, the Epyc comes really close (screenshots will be posted soon!).

Also waiting for MS_AT's confirmation on whether the binary contains AVX-512 instructions to further understand the Epyc's performance.

 

[Markfw]

Unfortunately, Ice Lake wins

Because:

The benchmark has some bug. It won't use more than 64 threads. I'll have to check the code tomorrow to see if there is a hard limit in there.

If Ice Lake is really using AVX-512, that might explain how 48 threads are able to beat 64 physical Zen 2 cores, though the Epyc puts up a fierce fight armed only with AVX2 and gets pretty close. Considering that the Ice Lake server cost $7000 and my Epyc cost not more than $1500, I would say the Epyc wins fair and square

However, if the compiler has "cheated" and replaced the AVX-512 code with AVX2 code, the only possible explanation for the Epyc's loss is higher frequency on the Ice Lake CPU (3.95 GHz vs. 2.91 GHz on the Epyc). But even then, with almost a GHz deficit, the Epyc comes really close (screenshots will be posted soon!).

Also waiting for MS_AT's confirmation on whether the binary contains AVX-512 instructions to further understand the Epyc's performance.

Rome has very limited avx-512 support. Zen 4 has quite a bit. Zen 5 kicks butt ! If I get time to load up my Turin, I will run it. Even at 2.3 ghz, it beats Zen 4s at 3.5 ghz in avx-512 stull.

 

[Hail The Brain Slug]

9800X3D. Seems kinda low?

 

[poke01]

7700X, it was on eco mode too. Also shocking its not that far from a 9800X3D

 

[igor_kavinski]

9800X3D. Seems kinda low?
View attachment 119246

Can't say unless someone posts their 9800X3D score for comparison.

BUT, you may lose up to 9% score if you don't run it with admin rights.

7-max may help a bit too but doesn't always work.

It's a pure floating point bench. Only taxes the FPU units themselves.

The current champ Ice Lake Xeon (until more users test) does 42 Mops/s per thread. Yours is doing 55.25

 

[igor_kavinski]

7700X, it was on eco mode too. Also shocking its not that far from a 9800X3D

Coz there is no huge data movement involved. This is a FPU stress test.

 

[Hail The Brain Slug]

Can't say unless someone posts their 9800X3D score for comparison.

BUT, you may lose up to 9% score if you don't run it with admin rights.

7-max may help a bit too but doesn't always work.

It's a pure floating point bench. Only taxes the FPU units themselves.

The current champ Ice Lake Xeon (until more users test) does 42 Mops/s per thread. Yours is doing 55.25

Disabling AVX512 made no difference to scores, which would seem to corroborate the assertion that the AVX512 instructions are getting compiled away.

Administrator did nothing for me.

7-Max didn't crash the application, at least. No change to scores either.

 

[Markfw]

Kinda low. It barely beats the icelake Xen !

 

[Markfw]

2nd run, a little better.

 

[igor_kavinski]

View attachment 119251

Kinda low. It barely beats the icelake Xen !

Which CPU is that? It's probably low coz no more than 64 threads are being used.

I was able to get a score higher than 1900 on the Zen 2 Epyc by disabling SMT.

 

[igor_kavinski]

Disabling AVX512 made no difference to scores, which would seem to corroborate the assertion that the AVX512 instructions are getting compiled away.

Crap!

Back to the drawing board

I have to agree that the compiler stripped out the AVX-512 instructions. Guess I need to push out v0.02a with a fix.

May need to have a dedicated AVX-512 binary coz can't afford to spend time on doing proper CPU feature detection.

 

[Hail The Brain Slug]

Crap!

Back to the drawing board

I have to agree that the compiler stripped out the AVX-512 instructions. Guess I need to push out v0.02a with a fix.

May need to have a dedicated AVX-512 binary coz can't afford to spend time on doing proper CPU feature detection.

A benchmark that uses different instruction sets based on what the CPU has available is also not a very good benchmark IMO. Consistency and comparability right out the window

Edit: Before someone bites my head off, I mean a benchmark like this where the intention is to test how quickly the CPU can do a specific operation. If different CPU's are doing different operations, what are you even trying to compare then? Since no real work is being done, it's not measuring how fast different CPU's can accomplish a greater task.

 

[igor_kavinski]

Edit: Before someone bites my head off, I mean a benchmark like this where the intention is to test how quickly the CPU can do a specific operation. If different CPU's are doing different operations, what are you even trying to compare then? Since no real work is being done, it's not measuring how fast different CPU's can accomplish a greater task.

You are right, in a way

 

[Markfw]

This is a little better with SMT disabled

well, after damn MS did an update on last boot, it won't pase !!!

But its 2308 !

 

[Markfw]

 

[igor_kavinski]

A benchmark that uses different instruction sets based on what the CPU has available is also not a very good benchmark IMO. Consistency and comparability right out the window

Well, yeah I agree that for proper comparison, both CPUs should use the same ISA extensions so the score tells us which chip is designed better.

 

[igor_kavinski]

View attachment 119261

You are officially the new champ!

Just to confirm, this is the Turin, right?

 

[Markfw]

Now if we can talk @Det0x to do it on his 9950x !

and to the last post, yes the Turin ES