Discussion Zen 7 speculation thread

[511]

And you know this how?

GB released an update for SME at the same time Apple released processor can't be co incidence let's see if it has APX/AVX 10.2 support on time or not

 

[poke01]

GB released an update for SME at the same time Apple released processor can't be co incidence let's see if it has APX/AVX 10.2 support on time or not

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

 

[poke01]

Or the fact that primate labs updated GB6.5 two weeks before Qualcomms keynote to support SME for WoA for X2 Elite benchmarks.

 

[511]

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

AVX -512 was introduced in SkylakeX 2017 for Xeons

 

[poke01]

AVX -512 was introduced in SkylakeX 2017 for Xeons

exactly so for select niche SKUs support was late. They added better support thanks to rocket lake and AVX512-FP16 support was added a few months after SPR launched, again due to mainstream SKUs not supporting these.


AVX512-FP16 is only for Xeons right now in GB and Zen6 should benefit next year since FP16 support is already there.

 

[okoroezenwa]

GB released an update for SME at the same time Apple released processor can't be co incidence let's see if it has APX/AVX 10.2 support on time or not

A month before.

 

[poke01]

This is how Zen6/7 instruction support may look like.


And boy do those math mul extensions accelerate multiple subtests

https://browser.geekbench.com/v6/cpu/14814281

 

[511]

This is how Zen6/7 instruction support may look like.
View attachment 133747

And boy do those math mul extensions accelerate multiple subtests

https://browser.geekbench.com/v6/cpu/14814281

this is gnr nvm

 

[Josh128]

I am not even sure how optimized GB is for AVX-512

All the Intel shills used to scream about it being the only reason Zen 4 beat Alder Lake in GB, but in reality turning off AVX-512 only made like 1%-2% difference in the score.

 

[511]

All the Intel shills used to scream about it being the only reason Zen 4 beat Alder Lake in GB, but in reality turning off AVX-512 only made like 1%-2% difference in the score.

1%

Gigabyte Technology Co., Ltd. B650M AORUS ELITE AX vs 9950X default (AVX512 off) - Geekbench

browser.geekbench.com

 

[OneEng2]

All the Intel shills used to scream about it being the only reason Zen 4 beat Alder Lake in GB, but in reality turning off AVX-512 only made like 1%-2% difference in the score.

Even if true, who cares? If Zen has something in it that makes it faster than Intel processors across a wide range of applications (even within a specific use model and market), the customer will see a reasonable difference ..... and that matters IMO.

Also, lets not get too cute here and forget that Intel put these instructions in first only to back them out because the implementation caused heat issues AND took up an inordinate amount of space. It wasn't because they weren't effective at adding performance.

Intel has decided to do a lesser approach to AVX 512 and use the remaining transistor budget for other things. This is IMO not a bad approach for them..... but that doesn't diminish the fact that AMD's AVX 512 implementation is amazing and provides substantial performance benefits.

 

[adroc_thurston]

6%
View attachment 133756

Gigabyte Technology Co., Ltd. B650M AORUS ELITE AX vs 9950X default (AVX512 off) - Geekbench

browser.geekbench.com

The 1t delta is literally 1%.

 

[511]

Even if true, who cares? If Zen has something in it that makes it faster than Intel processors across a wide range of applications (even within a specific use model and market), the customer will see a reasonable difference ..... and that matters IMO.

Not on windows and certainly not in video games we would need another 10 years for AVX-512 to be the baseline

 

[511]

The 1t delta is literally 1%.

you are correct my brain took 6% as ST for some dumb reason

 

[511]

this doesn't look correct to me wasn't it 2X difference?

 

[CouncilorIrissa]

View attachment 133758
this doesn't look correct to me wasn't it 2X difference?

the behaviour is reproducible on Windows as well, 10% in one subtest.

 

[Hitman928]

View attachment 133758
this doesn't look correct to me wasn't it 2X difference?

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.

 

[adroc_thurston]

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.

Pretty sure that was the compiler delta during one of GB's updates.

 

[Joe NYC]

exactly so for select niche SKUs support was late. They added better support thanks to rocket lake and AVX512-FP16 support was added a few months after SPR launched, again due to mainstream SKUs not supporting these.


AVX512-FP16 is only for Xeons right now in GB and Zen6 should benefit next year since FP16 support is already there.

Xeon CPU is not a niche processor.

GB banding over backwards to showcase Arm processors tells you everything about Geekbench.

 

[Joe NYC]

1%
View attachment 133756

Gigabyte Technology Co., Ltd. B650M AORUS ELITE AX vs 9950X default (AVX512 off) - Geekbench

browser.geekbench.com

In Geekbench 5, I believe the difference was greater than in GB 6.

 

[Markfw]

the behaviour is reproducible on Windows as well, 10% in one subtest.

Everybody is not even looking at code which is avx-512 heavy, as most primegrid applications are. I am sure there are commercial applications that are avx-512 heavy. It makes 30-40% difference in those cases. Phoronix saw 56%

Phoronix result here: https://www.phoronix.com/review/amd-zen5-avx-512-9950x/7


When taking the geometric mean of the 90 benchmarks used for this AVX-512 on/off comparison, the Zen 5 AVX-512 implementation with the Ryzen 9 9950X saw its performance go up by 56%

 

[511]

Everybody is not even looking at code which is avx-512 heavy, as most primegrid applications are. I am sure there are commercial applications that are avx-512 heavy. It makes 30-40% difference in those cases.

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

 

[Markfw]

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

what is GB ?

 

[Joe NYC]

1%
View attachment 133756

Gigabyte Technology Co., Ltd. B650M AORUS ELITE AX vs 9950X default (AVX512 off) - Geekbench

browser.geekbench.com

Here are 2 Geekbench 5 submissions that look like identical system. Both submissions by David Huang:

9950x AVX-512 off: 2425

9950X default (AVX512 off) - Geekbench

Benchmark results for a Gigabyte Technology Co., Ltd. B650M AORUS ELITE AX with an AMD Ryzen 9 9950X processor.

browser.geekbench.com

9950x AVX-512 on: 2652

Gigabyte Technology Co., Ltd. B650M AORUS ELITE AX - Geekbench

Benchmark results for a Gigabyte Technology Co., Ltd. B650M AORUS ELITE AX with an AMD Ryzen 9 9950X processor.

browser.geekbench.com

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

So, when you see you see 1%, it is after you have been visited by the Men In Black.

 

[Joe NYC]

Here are 2 Geekbench 5 submissions that look like identical system. Both submissions by David Huang:

9950x AVX-512 off: 2425

9950X default (AVX512 off) - Geekbench

Benchmark results for a Gigabyte Technology Co., Ltd. B650M AORUS ELITE AX with an AMD Ryzen 9 9950X processor.

browser.geekbench.com

9950x AVX-512 on: 2652

Gigabyte Technology Co., Ltd. B650M AORUS ELITE AX - Geekbench

Benchmark results for a Gigabyte Technology Co., Ltd. B650M AORUS ELITE AX with an AMD Ryzen 9 9950X processor.

browser.geekbench.com

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

So, when you see you see 1%, it is after you have been visited by the Men In Black.

View attachment 133770

MT:
AVX-512 off: 23,561
AVX-512 on: 26,093

Difference even greater: 10.6%