Discussion Zen 7 speculation thread

[511]

what is GB ?

Geekbench

 

[poke01]

It’s like 80% or something over Zen4 (or was it Zen4 over Zen 3?) and some people took the reason as being AVX512 support but that was obviously a false assumption.

That was just AVX512-VNNI being used for object detection

 

[Geddagod]

That's a 9.3% difference that Geekbench deceptively "erased", just to show x86 in worst light and Arm in the best light.

Does this not make ARM cores being very close to Zen 5 (and Apple outright ahead) in GB5 even worse then?

 

[poke01]

I even doubt proper AVX-512 support on GB ...

What do you mean proper support? do you want subtests like AES-XTS to come back?

Also not every work flow benefits from AVX512, AV1 encoding for example doesn’t benefit that much. It’s around a 10-20% improvement at best.

 

[Hitman928]

That was just AVX512-VNNI being used for object detection

So why does turning AVX512 off only decrease the object detection score by 10%? Is there a fallback to something else Zen5 added support for that’s not AVX512 that can give near the same boost?

 

[poke01]

So why does turning AVX512 off only decrease the object detection score by 10%? Is there a fallback to something else Zen5 added support for that’s not AVX512 that can give near the same boost?

yep Zen5 added normal AVX-VNNI

AMD Zen 5 Compiler Support Posted For GCC - Confirms New AVX Features & More - Phoronix

www.phoronix.com

 

[Jan Olšan]

Guess you didn’t see the GB5.1 update which improved AES-XTS for AVX-512 in 2019, when only Intel supported AVX-512.

That's 2 years late then.

Not sure about D.H.'s result but, people here tried to run GB6 on Zen 5 with AVX-512 disabled and enabled, and the difference was so close to margin of error... Even the +10% in two tests doesn't really sound like they used much of the opportunities to speed up stuff with 2x higher compute throughput. Generaly the set of benchmarks should have some fruit that AVX-512 can pick, if it can profit from stuff like SME. AVX-512 is more applicable to general tasks so you'd expect it to yield more overall, not really less.

 

[poke01]

That's 2 years late then.

Do you know when SVE support was added?

In 2023 with GB6.1, that is 5 years late when the first ARM cpu released with with SVE in 2018.

 

[Markfw]

Would you all please read my edited post here: https://forums.anandtech.com/threads/zen-7-speculation-thread.2630494/post-41534542

Its 90 different tests about avx512 and a 56% advantage using it turned on vs off.

 

[Hitman928]

yep Zen5 added normal AVX-VNNI

AMD Zen 5 Compiler Support Posted For GCC - Confirms New AVX Features & More - Phoronix

www.phoronix.com

Is GB compiled with that support though? I think the white paper only mentions it being supported with AVX-512 but it’s been a while and I don’t remember for sure. Also, doesn’t Zen 4 support VNNI? So, there shouldn’t really be that much difference between the two.

 

[poke01]

Also, people tried to run GB6 on Zen 5 with AVX-512 disabled and enabled, and the difference was so close to margin of error... It's clear that they didn't really use any of the opportunities to speed up something with 2x higher compute throughput - and the set of benchmarks should have some, if it can profit from stuff like SME.

Yeah cause Zen5 is boosted by AVX-VNNI already when AVX512 is turned off for the Object Detection test.

Zen3 to Zen4 had AVX512-VNNI support boosting Object detection by over 2.5x

Zen4 to Zen5 added AVX-VNNI which is a fall back when AVX512-VNNI is not available or turned off and that made it seem AVX512 does barely anything but AVX512 does increase it by 10%.




here is a 10900K vs 11900K and you can clearly see the speedups in photo library and Object detection

ASUS System Product Name vs To Be Filled By O.E.M. Z590 Phantom Gaming 4 - Geekbench

browser.geekbench.com

 

[Doug S]

Do you know when SVE support was added?

In 2023 with GB6.1, that is 5 years late when the first ARM cpu released with with SVE in 2018.

The only ARM CPU with SVE before was that was like Hitachi or some other Japanese company designed exclusively for HPC. Not much point in supporting it in Geekbench when no one will ever run it on that.

I assume when they did add SVE support it was when ARM announced cores that supported it. If there was some conspiracy where they wanted to show Apple in the best light they wouldn't have done that, because they never have and seems likely never will support SVE (Apple does support SSVE starting with M4, which is a different thing but it is a low performance implementation so unless something has changed with M5 was done solely because it is part of SME so they don't intend anyone to actually use it)

 

[StefanR5R]

It's OK that friends of Mac OS on AARCH 64 are fascinated with GeekBench. But why give this much room to this one benchmark in Zen 7 Speculation?

 

[poke01]

It's OK that friends of Mac OS on AARCH 64 are fascinated with GeekBench. But why give this much room to this one benchmark in Zen 7 Speculation?

because a lot of these extensions in SPR will be added to Zen7 which Geekbench already supports, which goes against x86 people thinking that there is barely any optimisations for x86 in this benchmark.

 

[poke01]

It’s funny when I show that Geekbench shows support and has optimisation for x86 extensions, the discussion moves to making fun of me.

 

[adroc_thurston]

It’s funny when I show that Geekbench shows support and has optimisation for x86 extensions, the discussion moves to making fun of me.

Ehhhh optimization is a bold word for very nominal levels of support in GB.

because a lot of these extensions in SPR will be added to Zen7 which Geekbench already supports, which goes against x86 people thinking that there is barely any optimisations for x86 in this benchmark.

Support doesn't mean it runs well.

 

[poke01]

Support doesn't mean it runs well.

True but having support is something and better than nothing. A 2.5x speed up over Zen3 to Zen4 is nothing to scoff at even if it’s not optimised properly.

Anyway I’ll move the talk of Geekbench optimisation to the Geekbench thread. This is going far from Zen7 now.

 

[adroc_thurston]

This is going far from Zen7 now.

It's another tock with more stuff.

 

[Hitman928]

Yeah cause Zen5 is boosted by AVX-VNNI already when AVX512 is turned off for the Object Detection test.

Zen3 to Zen4 had AVX512-VNNI support boosting Object detection by over 2.5x

Zen4 to Zen5 added AVX-VNNI which is a fall back when AVX512-VNNI is not available or turned off and that made it seem AVX512 does barely anything but AVX512 does increase it by 10%.




here is a 10900K vs 11900K and you can clearly see the speedups in photo library and Object detection


View attachment 133790

ASUS System Product Name vs To Be Filled By O.E.M. Z590 Phantom Gaming 4 - Geekbench

browser.geekbench.com

Thanks, that makes sense. The big difference I saw before in object detection between Zen4 and Zen5 must have been a version difference that included with and without support between the two or something.

 

[poke01]

It's another tock with more stuff.

I meant that the topic at hand is going away from Zen7 discussion. 😅

 

[LightningDust]

The only ARM CPU with SVE before was that was like Hitachi or some other Japanese company designed exclusively for HPC. Not much point in supporting it in Geekbench when no one will ever run it on that.

Fujitsu A64FX. Hitachi hasn't made a CPU inhouse since the AP8800E, a mainframe chip around 2013.

Also, Neoverse V1 was announced 2021, though shipping silicon was 2022.

 

[Joe NYC]

Does this not make ARM cores being very close to Zen 5 (and Apple outright ahead) in GB5 even worse then?

If you depreciate AVX-512 to the point that they lose 10% of the GB score and add SME support that ads ~10% to Arm, effectively create a 20% swing, GB can no longer be called a "benchmark", but a PR Arm of Arm.

Something to keep in mind when the Arm horde point to GB "benchmark" to always call BS on it.

If they did not subtract 10% from AVX-512 and added SME, that would be explained a reasonable but doing both (subtract from x86 and add to Arm) = rigged.

 

[poke01]

If you depreciate AVX-512 to the point that they lose 10% of the GB score and add SME support that ads ~10% to Arm, effectively create a 20% swing, GB can no longer be called a "benchmark", but a PR Arm of Arm.

Something to keep in mind when the Arm horde point to GB "benchmark" to always call BS on it.

If they did not subtract 10% from AVX-512 and added SME, that would be explained a reasonable but doing both (subtract from x86 and add to Arm) = rigged.

What are you taking about?

Geekbench 5.5 and Geekbench 6.5 use different subtests. Geekbench 6 is more targeted towards ML.

So it’s okay if Geekbench boosts an x86 processor in GB5 by 2x and doesn’t boost ARM scores.

They didn’t subtract anything for x86, in fact it’s even greater thanks to AVX512-VNNI

Without AVX512-VNNI and AVX-VNNI support total scores for Zen5 and Lion Cove would be lower by 10% as well in GB6.

if anything SVE2 is not supported in Geekbench despite Android phones supporting it since 2022, so please don’t do argument that Geekbench is ARM PR when Geekbench supports niche x86 extensions like AVX512-FP16 which is supported only on GNR and SPR.

Get started with Scalable Vector Extension 2 (SVE2) on Android: Implement vector operations

This is an introductory topic for software developers interested in learning how to use the Scalable Vector Extension 2 (SVE2) on Arm powered mobile devices running Android.

learn.arm.com

 

[Joe NYC]

Geekbench 5.5 and Geekbench 6.5 use different subtests. Geekbench 6 is more targeted towards ML.

Yes, they took out / depreciated those that favored AVX-512 and added those that favor SME.

It's called rigging the results.

 

[poke01]

Yes, they took out / depreciated those that favored AVX-512 and added those that favor SME.

It's called rigging the results.

Are you being ignorant on purpose?

Object detection in Geekbench 6 favoured AVX512-VNNI by a factor of 2.5x way before SME even came into the picture.