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

[DisEnchantment]

Speculate at will

 

[Tup3x]

So only 10% uplift on windows? Excuse me but wtf

Geekbench doesn't necessarily tell the whole story.

 

[igor_kavinski]

So only 10% uplift on windows? Excuse me but wtf

You only need to be concerned if Geekbench's actual code for the subtests is being used in any or most commercial software.

Geekbench score is only a reflection of how well the CPU runs Geekbench code. Nothing else. It's an approximation of real world code but nobody spends AS MUCH time on tweaking their code as this Geekbench developer dude. Most commercial real world code is either super optimized or super turdy, with the majority in the latter class.

In like maybe 99% cases, if a software is fulfilling its purpose, the developer will NEVER again touch its code to tweak it because a codebase is a virtual bomb that can explode with the slightest change and that explosive probability is proportional to the size of the codebase and number of developers involved.

 

[yuri69]

irrespective of frequency

The reported frequency results are most likely mixed with the E-cores. This makes the frequency spread.

 

[Hitman928]

What is funny is that most of the public Strix GB6 scores (that I could discover online) land around that range irrespective of frequency

View attachment 102246

GB reported frequency for Strix has been obviously unreliable from the first samples. I was hoping that with the more retail samples showing up, that would get fixed, but it seems to still be an issue.

 

[eek2121]

That would be called a dGPU - the most modern ones are able to be used for this purpose, and they can be used for gaming, too!

Not really, NPUs lack dedicated TMUs and a bunch of other stuff.

You only need to be concerned if Geekbench's actual code for the subtests is being used in any or most commercial software.

Geekbench score is only a reflection of how well the CPU runs Geekbench code. Nothing else. It's an approximation of real world code but nobody spends AS MUCH time on tweaking their code as this Geekbench developer dude. Most commercial real world code is either super optimized or super turdy, with the majority in the latter class.

In like maybe 99% cases, if a software is fulfilling its purpose, the developer will NEVER again touch its code to tweak it because a codebase is a virtual bomb that can explode with the slightest change and that explosive probability is proportional to the size of the codebase and number of developers involved.

Geekbench does use many libraries that are used in both commercial and open source applications.

Whitepaper is here: https://www.geekbench.com/doc/geekbench6-benchmark-internals.pdf

 

[eek2121]

GB reported frequency for Strix has been obviously unreliable from the first samples. I was hoping that with the more retail samples showing up, that would get fixed, but it seems to still be an issue.

It is likely due to the frequencies of the Zen5c cores being detected instead of the “big” cores.

 

[igor_kavinski]

I was hoping that with the more retail samples showing up, that would get fixed, but it seems to still be an issue.

Maybe it's a feature? Press release from AMD on 26th July: "AMD implements the world's first CPU based on quantum frequency fluctuation made possible by the mystery of quantum tunnelling!".

 

[Hitman928]

It is likely due to the frequencies of the Zen5c cores being detected instead of the “big” cores.

That could be a big part of it, but that wouldn’t explain all of it.

 

[CouncilorIrissa]

It is likely due to the frequencies of the Zen5c cores being detected instead of the “big” cores.

Some of the reported frequencies are quite a bit higher than what Huang found the C-cores to clock at (although on an ES; 3.3 GHz).

 

[adroc_thurston]

It is likely due to the frequencies of the Zen5c cores being detected instead of the “big” cores.

Dense ain't running 4GHz on Strix.

 

[eek2121]

Dense ain't running 4GHz on Strix.

I was referring to all the results with low-mid 3ghz.

 

[leoneazzurro]

Not really, NPUs lack dedicated TMUs and a bunch of other stuff.

What I mean is, if you need to reach the 40+ TOPS for the Copilot certification and you need to use a PCIE device for it, a dGPU would be the most sensible option, as it serves to multiple purposes.

 

[eek2121]

What I mean is, if you need to reach the 40+ TOPS for the Copilot certification and you need to use a PCIE device for it, a dGPU would be the most sensible option, as it serves to multiple purposes.

It is going to come down to cost, efficiency, and performance scaling. In theory, an NPU should cost less and use less power.

 

[leoneazzurro]

It is going to come down to cost, efficiency, and performance scaling. In theory, an NPU should cost less and use less power.

Well, if you plan to have a dGPU you don't need a TPU at all, except corner cases whereyou need LOT of TOPS and/or you don't want to tax the GPU at the same time.

 

[igor_kavinski]

Geekbench does use many libraries that are used in both commercial and open source applications.

Replied here: https://forums.anandtech.com/thread...d-against-core-i7-12700.2610597/post-41243831

Because I sense you will disagree. A LOT

 

[Goop_reformed]

Geekbench doesn't necessarily tell the whole story.

All I care about at this point is the gaming performance and average st uplift, and the leaks just doesn't sit right with me

 

[Fjodor2001]

Well, if you plan to have a dGPU you don't need a TPU at all, except corner cases whereyou need LOT of TOPS and/or you don't want to tax the GPU at the same time.

The discussion was started regarding cases where no dGPU is used. I.e. how to fulfill the 40 TOPS requirement then.

If an NPU is included in the CPU like on Arrow Lake DT then the problem is solved. But on Zen5 DT it needs to be solved in some other way, unless it will include an NPU too (which we’ve not heard anyhing about so far). Possibly there can be PCIe NPUs that can be bought as add-on. Just like you can buy TPMs as add-on for PCs that don’t have that.

Should be cheaper to buy an NPU than dGPU, if the intention only is to reach 40 TOPS and there is no need for dGPU and the additional functionality that it provides. For the reasons eek2121 mentioned.

 

[igor_kavinski]

If an NPU is included in the CPU like on Arrow Lake DT then the problem is solved.

Do we know the TOPS for that?

 

[Doug S]

Yeah that’s on Linux. Linux typically scores 6-10% higher.

Has there ever been a decent explanation as to why? Is it running significantly different code on Windows vs Linux? Using suboptimal compiler options on Windows? Making "libc" calls that are translated through some intermediate library? Perhaps Windows is more aggressive in trying to keep the CPU in a lower power / lower frequency state?

 

[adroc_thurston]

I was referring to all the results with low-mid 3ghz.

These never existed, the slowest run is like 3.7.

 

[igor_kavinski]

Has there ever been a decent explanation as to why?

Linux Kernel isn't a bloated mess. Certain distros even run Win32 games faster than Windows.

Three gaming-focused Linux operating systems beat Windows 11 in gaming benchmarks

All three Linux distros managed to beat Windows 11 while using Valve's Proton compatibility layer.

www.tomshardware.com

When dumb HR folks measure programmer productivity in lines of code (loc), there's bound to be inventions of artificial complexity in code to inflate the loc. And then future programmers with common sense won't touch that codebase because pointing out flaws or justification for a complete rewrite of even a simple oft used function would be seen as disrespect of the senior programmers.

 

[igor_kavinski]

Should be cheaper to buy an NPU than dGPU

Only if there is enough competition. Without competition, even snake oil fetches a good price.

 

[leoneazzurro]

The discussion was started regarding cases where no dGPU is used. I.e. how to fulfill the 40 TOPS requirement then.

If an NPU is included in the CPU like on Arrow Lake DT then the problem is solved. But on Zen5 DT it needs to be solved in some other way, unless it will include an NPU too (which we’ve not heard anyhing about so far). Possibly there can be PCIe NPUs that can be bought as add-on. Just like you can buy TPMs as add-on for PCs that don’t have that.

Should be cheaper to buy an NPU than dGPU, if the intention only is to reach 40 TOPS and there is no need for dGPU and the additional functionality that it provides. For the reasons eek2121 mentioned.

Are there 40+ TOPS PCI-E TPUs costing less than a low-end modern dGPU? Because the ones I see around are quite expensive.

 

[Fjodor2001]

Are there 40+ TOPS PCI-E TPUs costing less than a low-end modern dGPU? Because the ones I see around are quite expensive.

I haven’t seen any NPUs in the 40 TOPS range. Only models with many more TOPS, which thus of course are more expensive than a low-end dGPU.

The reason is likely that there hasn’t been any need for NPUs in the 40 TOPS range until now. Regarding potential market for such NPUs, I guess they are not only for Zen5, but older CPUs too such as Zen4 and earlier, and corresponding Intel CPUs too. Basically for all CPUs without included NPU, that you’d like to upgrade to fulfill the Win12 AI PC requirements.

 

[whoshere]

Has there ever been a decent explanation as to why? Is it running significantly different code on Windows vs Linux? Using suboptimal compiler options on Windows? Making "libc" calls that are translated through some intermediate library? Perhaps Windows is more aggressive in trying to keep the CPU in a lower power / lower frequency state?

I'm curious myself and I even asked John Poole this question:

Linux vs. Windows results / Geekbench / Discussion Area - Primate Labs Support

support.primatelabs.com

Speaking of compilers:

For GB6.3.0:

• The Linux version could be using CLANG 16.0.0 and GCC Ubuntu 7.5.0-3ubuntu1~18.04.
• The Windows version could be using MSVC 10.0. But then it contains references to GCC and LLVM, so I've no clue. Looks like it contains object files made by different compilers.