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

[DisEnchantment]

Speculate at will

 

[DrMrLordX]

That's just exaggeration, they just couldn't install Nvidia drivers on the 14900K machine(s), install always failed towards the end.

Oof. Well at least AMD recalled their Zen5 CPUs so hopefully that won't happen on Granite Ridge!!!
!!
!

 

[Josh128]

9700X 9900X CB R23

Looks like we can kiss +17% R23 ST goodbye.

https://twitter.com/x/status/1817010872864375004

 

[gdansk]

It's 18% higher than my 65W 7700.
In any case this is margin-of-cooler kind of stuff here. Exhilarating.

 

[biostud]

It just goes to show you can't please everyone. I remember reading nothing but "AMD shouldn't ship these chips so far out of their efficiency range!" and now it seems the majority of the sentiment is "Why would they use such a low TDP if it's only XX% faster than Zen4?! They're leaving so much performance on the table!"

Yup, I've said exactly the same thing.
I actually think AMD prefer the Zen5 way, but with Zen4 they had to crank it up to stay competitive and, the efficiency was still way better than Intels offerings. With Intels next generation chips they want to stay ahead in efficiency as well, and let PBO handle o/c for those who want more performance at the cost of efficiency. Also because every gamer will get the X3D version anyway, and the non 3D version are targeting work PC, where a little performance decrease doesn't matter, if it can reduce power/cooling needs.

 

[ErichZannMusician]

9700X 9900X CB R23

Looks like we can kiss +17% R23 ST goodbye.

https://twitter.com/x/status/1817010872864375004

The first picture is the clearest evidence yet of the Zen 5 front-end cvcking AMD's SMT yield.

 

[gdansk]

The first picture is the clearest evidence yet of the Zen 5 front-end cvcking AMD's SMT yield.

Not the TDP?

 

[ErichZannMusician]

Not the TDP?

No, because there would be marginal difference between the MP ratio of the 65 W 7700 vs the 65 W 9700X:

AMD Ryzen 5 7600, Ryzen 7 7700 and Ryzen 9 7900 Review

Ryzen 7000 CPUs have not been a slam dunk for AMD so far. Even though they are quite speedy, the move to the AM5 platform has proved...

www.techspot.com

7700 MT ratio = 9.58
9700X MT ratio = 9.44

Granted it is small, but the theory is supported by the data.

 

[gdansk]

No, because there would be marginal difference between the MP ratio of the 65 W 7700 vs the 65 W 9700X:

AMD Ryzen 5 7600, Ryzen 7 7700 and Ryzen 9 7900 Review

Ryzen 7000 CPUs have not been a slam dunk for AMD so far. Even though they are quite speedy, the move to the AM5 platform has proved...

www.techspot.com

7700 MT ratio = 9.58
9700X MT ratio = 9.44

Granted it is small, but the theory is supported by the data.

Isn't it hard to say without normalizing for clock rates?

 

[ErichZannMusician]

Isn't it hard to say without normalizing for clock rates?

True, but the rate at which anecdotes are piling up increasingly point to that being indeed the case.

Remember the recent tweet from a laptop reviewer about weirdness in Cyberpunk 2077 performance?

 

[gdansk]

True, but the rate at which anecdotes are piling up increasingly point to that being indeed the case.

Remember the recent tweet from a laptop reviewer about weirdness in Cyberpunk 2077 performance?

Hadn't heard of it personally but if measuring performance on a dual CCX laptop with separate core types there is a lot of room for scheduling tomfoolery, clock rate differences and so on.

 

[ErichZannMusician]

Hadn't heard of it personally but if measuring performance on a dual CCX laptop with separate core types there is a lot of room for scheduling tomfoolery, clock rate differences and so on.

If dual - CCX was the sole reason, why specifically mention Cyberpunk?

SMT on AMD was broken since launch of Cyberpunk, and wasn't fixed till the 2.0 patch almost 3 years after release.

 

[PJVol]

I would actually lump CPU-Z bench

What about CB R23? I told you, it's a dull appearance, nothingburger, Zen4+

9700X MT +8% ST+13% (zen5 > zen4)
9900X MT +10% ST+10%

7700X +29% +25% (zen4 > zen3)
7900X +37% +24%

 

[Philste]

9700X MT +8% ST+13% (zen5 > zen4)
9900X MT +10% ST+10%

MT fits to AMDs claims, ST however... it should be 17% for 9900X and 19% for 9700X

 

[igor_kavinski]

I suspect that Zen 5 performance gains in most existing non-AVX-512 software will be meager until software is optimized and recompiled for Zen 5. I've said it before many times. Zen 5 is a FOSS powering design where compiling binaries before deployment or serious usage isn't an alien concept.

The crazy trajectory that Windows 11 seems to be on and the fact that I don't like the UI changes Microsoft has made to it in comparison to their far better Windows 10, makes me confident that I'll switch to mostly Linux based computing in the future and so Zen 5 seems like a very attractive option to me that will give me years of loyal service with incremental improvement in software performance.

I'll wait at least a year post launch for a Phoronix anniversary article on Zen 5 performance before calling it a less successful design than Zen 3 or Zen 4 were over their predecessors.

 

[Philste]

Does Strix Review Embargo drop today? Because sales are starting tomorrow.

 

[Nothingness]

I suspect that Zen 5 performance gains in most existing non-AVX-512 software will be meager until software is optimized and recompiled for Zen 5. I've said it before many times. Zen 5 is a FOSS powering design where compiling binaries before deployment or serious usage isn't an alien concept.

The crazy trajectory that Windows 11 seems to be on and the fact that I don't like the UI changes Microsoft has made to it in comparison to their far better Windows 10, makes me confident that I'll switch to mostly Linux based computing in the future and so Zen 5 seems like a very attractive option to me that will give me years of loyal service with incremental improvement in software performance.

I'll wait at least a year post launch for a Phoronix anniversary article on Zen 5 performance before calling it a less successful design than Zen 3 or Zen 4 were over their predecessors.

And you'd have to recompile all your Linux distro from scratch with the proper Zen5 compiler flags. I have never done that as I never felt the need during my 30 years of using Linux because my new CPU was always performing better by default than the previous one.

If a CPU requires recompilation of existing applications to perform well, I call that a failure. Note "perform well" is different from recompiling to extract more performance than the default on existing binaries, which I can understand and which I'd like to see benchmarked.

The last few times I tested native compiler flags on the various CPUs I used, the benefit was in the noise (except of course for new instructions, but that doesn't imply the use of specific instruction scheduling rules), so I've stopped wasting my time on that. If Zen5 requires compiler tuning and recompilation to perform well, I'd be extremely disappointed (I've seen no proof of that yet; sure there might be some outliers, but I think in general the perf uplift looks good).

TLDR: A modern successful CPU should not require recompilation of applications to perform well.

 

[igor_kavinski]

TLDR: A modern successful CPU should not require recompilation of applications to perform well.

AMD Ryzen 9 7950X "Zen 4" Rocks On Intel's Clear Linux - Phoronix

www.phoronix.com

Even Zen 4 sees a performance improvement when running on an Intel optimized (and thus recompiled) distro.

 

[Nothingness]

AMD Ryzen 9 7950X "Zen 4" Rocks On Intel's Clear Linux - Phoronix

www.phoronix.com

Even Zen 4 sees a performance improvement when running on an Intel optimized (and thus recompiled) distro.

View attachment 103989

That likely is the use of new instructions. Do you really think Intel used AMD specific compiler flags to build their Clear Linux?

Edit: flags show that Clear Linux tuning is for Skylake. So yeah definitely no AMD tuning. You're basically proving my point: no uarch specific tuning was needed for Zen4 and I hope it will be the same for Zen5.

 

[Kryohi]

And you'd have to recompile all your Linux distro from scratch with the proper Zen5 compiler flags.

Not necessarily. There are many distros nowadays that ship optimized x86_64-v4 (AVX512) binaries, and some are going further, having repositories for packages compiled for zen4 and zen5 uarchs (CachyOS). I don't know how extensive these repos are, but it should be feasible to include most common and not-so-common packages beyond the system ones.

A modern successful CPU should not require recompilation of applications to perform well

Definitely, but what's the problem if a part of the performance increase can only be obtained after recompilation? It's been the same for many years tbh, how many years did "normal users" had to wait before AVX2 became widely used?

 

[Nothingness]

Not necessarily. There are many distros nowadays that ship optimized x86_64-v4 (AVX512) binaries, and some are going further, having repositories for packages compiled for zen4 and zen5 uarchs (CachyOS). I don't know how extensive these repos are, but it should be feasible to include most common and not-so-common packages beyond the system ones.


Definitely, but what's the problem if a part of the performance increase can only be obtained after recompilation? It's been the same for many years tbh, how many years did "normal users" had to wait before AVX2 became widely used?

I certainly agree with you. But this is different from having to do specific tuning for a uarch (rather than targetting new instructions). And that's the point I'm arguing about 😀

As far as the specific repos for zen4 go, I wonder what the benefit of fine tuning for that specific uarch vs only using new instructions supported by zen4 is. If it's less than 2/3% that's what I call noise (and I've seen examples where using the compiler native flag was producing slightly slower executables).

Edit: sorry I forgot to address a specific point. I have no problem with recompiling to extract the last bit of performance. The problem is if I have to do it to gain anything from a new CPU. That shows the CPU has issues.

 

[poke01]

CachyOS is a niche of a niche, it’s an enthusiast OS with a great community. It will be great OS to move to if MS keeps on filling Windows with crapware.
——
Here is the thing tho, CPUs are general purpose and they should work great OOTB. Unlike GPUs which require drivers and optimisations and continuous improvement , CPUs need to just work and performance should be there day 1 for 99.99% of tasks. Zen 5 lays the ground work for future Zen architectures.

We know a good architecture when we see one because it stands out not only in benchmarks but in real world tests.

 

[fastandfurious6]

when is review out!!!!!!!!!!!!!!!!!!!!

 

[Josh128]

I suspect that Zen 5 performance gains in most existing non-AVX-512 software will be meager until software is optimized and recompiled for Zen 5. I've said it before many times. Zen 5 is a FOSS powering design where compiling binaries before deployment or serious usage isn't an alien concept.

The crazy trajectory that Windows 11 seems to be on and the fact that I don't like the UI changes Microsoft has made to it in comparison to their far better Windows 10, makes me confident that I'll switch to mostly Linux based computing in the future and so Zen 5 seems like a very attractive option to me that will give me years of loyal service with incremental improvement in software performance.

I'll wait at least a year post launch for a Phoronix anniversary article on Zen 5 performance before calling it a less successful design than Zen 3 or Zen 4 were over their predecessors.

Igor, assuming the ES user that provided the leaks earlier still has it, is there any way it would be possible to get him to run a locked 4 or 5 GHz ST run to further analyze the apparent ST discrepancy vs AMDs +17% R23 IPC claims that we are seeing in these R23 leaks?

 

[PJVol]

MT fits to AMDs claims

What claims? Did I miss something?

 

[Philste]

What claims? Did I miss something?

They claimed a 11-22% MT uplift in Blender. (11% for 9700X, 16% for 9600X, 17% for 9900X and 22% for 9950X). Blender has 23% IPC according to them. So for 17% IPC R23 you get like 8-16% uplift. 8% for 9700X, ~12% for 9600X and 9900X and 16% for 9950X.