Discussion Zen 5 Speculation (EPYC Turin and Strix Point/Granite Ridge - Ryzen 9000)

[DisEnchantment]

Speculate at will

 

[Hitman928]

My last link post 11544 is.

See post 11478 for why your links and comments aren’t relevant.

 

[igor_kavinski]

Geekerwan doesn't.

How can you be so positive about him?

 

[SarahKerrigan]

How can you be so positive about him?

Because they say it outright that they use gcc.

I'm not going to indulge conspiracy-theory garbage. "Everyone doing SPEC runs is secretly using icc when they say gcc!" is the kind of nonsense that isn't worth touching.

 

[Abwx]

Wasn't the original point of contention a modified POVRAY binary used by some Intel slurpee slurping website that disabled AVX2 support for Ryzen? Where did SPEC enter into this conversation?

Because a guy called Geekerwan posted an excel sheet showing the same difference in Povray within SPEC than in povray as used by Computerbase (wich are proved to be flawed in respect of ISA support) and his numbers were taken at face value even if it s not an official SPEC submission.

 

[SarahKerrigan]

Because a guy called Geekerwan posted an excel sheet showing the same difference in Povray within SPEC than in povray as used by Computerbase, wich are proved to be flawed in respect of ISA suppoprt, and his numbers were taken at face value even if it s not an official SPEC submission.

Anandtech showed a similar percent difference between Intel's and AMD's povray SPECfp numbers in their 7950X review. Are they in on your conspiracy too?

 

[igor_kavinski]

I'm not going to indulge conspiracy-theory garbage. "Everyone doing SPEC runs is secretly using icc when they say gcc!" is the kind of nonsense that isn't worth touching.

OK boss

 

[eek2121]

I am of the opinion that a good benchmark should test both “non-accelerated” int/fp performance and accelerated performance (AVX2, SSE, AVX512, etc)

My reasoning behind this because a chip can be a dinosaur at general compute, but have ridiculous accelerators in certain scenarios (see: Apple M4) that may not always apply.

It is too bad there isn’t any benchmark that does this. (yet! 😉)

Rather than butchering GB5, they should have done this instead.

100% sure this screenshot is fake. Looks like they just messed around with the CPU ID on a Zen 4 part.

I don’t do much validation, I just share them! 🤣

 

[igor_kavinski]

Anandtech showed a similar percent difference between Intel's and AMD's povray SPECfp numbers in their 7950X review. Are they in on your conspiracy too?

Hint: he's relentless. Give up!

 

[Saylick]

So Zen 5 30% IPC when? My feeble brain can't comprehend all this icc/gcc Spec this, Spec that mumbo jumbo.

 

[eek2121]

Nobody but Intel compiles with ICC. Anandtech doesn't. Geekerwan doesn't. Running it on Ryzen would likely be a little slower but not significantly - ICC generates decent code for AMD CPUs these days (but is itself increasingly irrelevant because gcc and clang have gotten really good.)

It has been a long time since I have looked into ICC, but at one point ICC actually churned out faster binaries for AMD than for Intel.

 

[igor_kavinski]

So Zen 5 30% IPC when?

When you turn it upside down.

When it's the right way up, it will reach max IPC of 62%, fortifying my position as the King of Speculation on these forums. I say that with the utmost humility.

P.S., bet you didn't know that Zen 5 learns to predict better the longer it runs. So those who refuse to reboot their PCs and maintain maximum uptime will be rewarded handsomely for their stubbornness.

 

[Timorous]

So Zen 5 30% IPC when? My feeble brain can't comprehend all this icc/gcc Spec this, Spec that mumbo jumbo.

@igor_kavinski already confirmed +99% so just waiting for the results next week now.

*ducks*

 

[Abwx]

Anandtech showed a similar percent difference between Intel's and AMD's povray SPECfp numbers in their 7950X review. Are they in on your conspiracy too?

You ll notice that these are estimations because they dont use a fixed frequency, also you surely didnt notice that they measured 7.25% difference.

On the other hand geekerwan "measure" 13.5%, wich you said is similar, if almost 2x the difference is similar and not suspicipous for you then i dont know what will.

 

[SarahKerrigan]

You ll notice that these are estimations because they dont use a fixed frequency, also you surely didnt notice that they measured 7.25% difference.

On the other hand geekerwan "measure" 13.5%, wich you said is not suspicious, if almost 2x the difference is similar and not suspicipous for you then i dont what will.

They're estimations because all results that don't go through the official SPEC submission process have to be marked as estimates.

Tell you what. I'm a SPEC licensee. If anyone wants to set me up with remote access to the relevant systems, I would be happy to run the relevant subtests myself. (But then, maybe I'm part of the conspiracy too...)

 

[Abwx]

They're estimations because all results that don't go through the official SPEC submission process have to be marked as estimates.

Tell you what. I'm a SPEC licensee. If anyone wants to set me up with remote access to the relevant systems, I would be happy to run the relevant subtests myself. (But then, maybe I'm part of the conspiracy too...)

That you are a SPEC licencee doesnt change the fact that you estimated that 7.25% and 13.5% are similar, we re not talking of 1.45% to 2.7%, wich would still be within a margin error.

And dont brand your licencee status as some cover, if geekerwan has greatly exagerated this number he could have done the same for others benches, truth is that it went undetected while the discrepancy was on front of your eyes, and you even used it as a "prove" that you were on point.

 

[Hitman928]

That you are a SPEC licencee doesnt change the fact that you estimated that 7.25% and 13.5% are similar, we re not talking of 1.45% to 2.7%, wich would still be within a margin error.

And dont brand your licencee status as some cover, if geekerwan has greatly exagerated this number he could have done the same for others benches, truth is that it went undetected while the discrepancy was on front of your eyes, and you even used it as a "prove" that you were on point.

13.5% versus 7.25% from 2 completely independent reviewers is a similar result. If you have a reason to question the results with actually relevant data, please share, otherwise you are just claiming that there's an issue because you don't like the results.

 

[SarahKerrigan]

That you are a SPEC licencee doesnt change the fact that you estimated that 7.25% and 13.5% are similar, we re not talking of 1.45% to 2.7%, wich would still be within a margin error.

And dont brand your licencee status as some cover, if geekerwan has greatly exagerated this number he could have done the same for others benches, truth is that it went undetected while the discrepancy was on front of your eyes, and you even used it as a "prove" that you were on point.

My brother in Christ, those are similar. That's not much of a discrepancy. Numbers shift around a fair bit depending on system configuration, compiler, etc. I'm still not inclined to consider a 6% delta between two different users suspicious and I don't know why you're trying to gotcha me about it.

As an example, between gcc 7.3 and 9.1 , in my own testing, on the same Rome machine, 631.deepsjeng_s improved by 18%. Other subtests actually regressed. x264_s got 5% slower.

 

[SarahKerrigan]

13.5% versus 7.25% from 2 completely independent reviewers is a similar result. If you have a reason to question the results with actually relevant data, please share, otherwise you are just claiming that there's an issue because you don't like the results.

Two reviewers with completely different compilers! Geekerwan was gcc, Anandtech was clang! Tbh I consider those numbers to be remarkably similar, under the circumstances.

 

[igor_kavinski]

Tell you what. I'm a SPEC licensee. If anyone wants to set me up with remote access to the relevant systems, I would be happy to run the relevant subtests myself.

At your behest, my lady! My 12700K beckons you to honor it with your smooth fingertipped keystrokes so it can know pleasure no man can fathom!

 

[Thunder 57]

This just in! Zen 5 slower than Raptor Lake!

Of course there is speculation if that screenshot is even real. Isn't this the same source that said IF and memory saw nice gains?

 

[igor_kavinski]

This just in! Zen 5 slower than Raptor Lake!

Thunder, ThunDER! THUNDERRRR!!! HOOOOOOOOO !!!

Quick, to the batmobile, before all RPL stock runs out!

 

[Abwx]

My brother in Christ, those are similar. That's not much of a discrepancy. Numbers shift around a fair bit depending on system configuration, compiler, etc. I'm still not inclined to consider a 6% delta between two different users suspicious and I don't know why you're trying to gotcha me about it.

As an example, between gcc 7.3 and 9.1 , in my own testing, on the same Rome machine, 631.deepsjeng_s improved by 18%. Other subtests actually regressed. x264_s got 5% slower.

Rest assured that there s no gotcha from my part, that being said there could be differences that large in INT based code, but not so much in FP like Povray where most cycles are dumped on arithmetic computations rather than data manipulations.

But let say that it s comparable if that please you.

 

[igor_kavinski]

Two reviewers with completely different compilers!

Two reviewers. Two compilers. Same agenda. Think like your adversary to understand his position!

 

[Kepler_L2]

This just in! Zen 5 slower than Raptor Lake!

Of course there is speculation if that screenshot is even real. Isn't this the same source that said IF and memory saw nice gains?

It's 100% fake

 

[SarahKerrigan]

At your behest, my lady! My 12700K beckons you to honor it with your smooth fingertipped keystrokes so it can know pleasure no man can fathom!

You made it weird

Rest assured that there s no gotcha from my part, that being said there could be differences that large in INT based code, but not so much in FP like Povray where most cycles are dumped on arithmetic computations rather than data manipulations.

But let say that it s comparable if that please you.

So, a couple points - first, if you read the Limaye-Adegbija paper (which I highly recommend) you'll see that Povray is actually branchier and memory-ops-ier than you'd probably expect, though they tend to be fairly easily predictable and fit well in caches. But second, sometimes arithmetic-heavy code shows the largest delta between compilers because successfully vectorizing one loop can make a big difference in end performance.