[CPU only] Apple M1/2(Max/Ultra) (TSMC 5nm) vs AMD Zen 4 (TSMC 5nm) - Technical Analysis

[Xiao_Xi]

It doesn't. The geomean SPEC2017 scores for 7950X (full TDP) are 92.5 and 100.9 for int and fp tests, respectively. For M1 Max (also from anandtech benchmarks) its 53.3 and 81.07. Double that for the Ultra. So Zen4 is about the same performance in int tests and considerably slower for fp tests. This is also reflected by Geekbench where 7950X and Ultra score about the same. Zen4 CPUs are of course faster in single core in almost every test. Of course, 7950X uses much more power.

What is the penalty for running SPEC2017 in WSL?

This is particularly obvious in floating point SPEC, where the unrestricted 7950X is only 25% faster than M1 Max despite using much more power.

How much does 7950X score in SPEC2017 at 65W? The 7950X can achieve at 65W 80% of the score in stock conditions.

Cinebench performance penalty for ARM CPUs is quite huge - which is no surprise as Embree is open source and just a quick glance reveals the issue.

Benchmarks are good or not depending on how well they reflect the software situation. If software similar to Cinebench is optimized for x86 and not for ARM, Cinebench results are fair and should be taken into account. By the way, it's up to Apple to fix Embree's performance on their hardware and make sure the next version of Cinebench includes an ARM-optimized version.

 

[leman]

What is the penalty for running SPEC2017 in WSL?

What’s WSL?

How much does 7950X score in SPEC2017 at 65W? The 7950X can achieve at 65W 80% of the score in stock conditions.

No idea. If you want to assume the same 80% performance retention then it’s going to be 50% faster in int tests and same in FP tests as the M1 Max.

 

[Xiao_Xi]

What’s WSL?

Windows Subsystem for Linux

We run the tests in a harness built through Windows Subsystem for Linux, developed by Andrei Frumusanu. WSL has some odd quirks, with one test not running due to a WSL fixed stack size, but for like-for-like testing it is good enough. Because our scores aren’t official submissions, as per SPEC guidelines we have to declare them as internal estimates on our part.

AMD Zen 4 Ryzen 9 7950X and Ryzen 5 7600X Review: Retaking The High-End

www.anandtech.com

 

[leman]

Windows Subsystem for Linux

AMD Zen 4 Ryzen 9 7950X and Ryzen 5 7600X Review: Retaking The High-End

www.anandtech.com

Ah ok, then zero overhead.

 

[Gerdi]

What is the penalty for running SPEC2017 in WSL?

No penalty for CPU benchmarks.

Benchmarks are good or not depending on how well they reflect the software situation. If software similar to Cinebench is optimized for x86 and not for ARM, Cinebench results are fair and should be taken into account. By the way, it's up to Apple to fix Embree's performance on their hardware and make sure the next version of Cinebench includes an ARM-optimized version.

That depends what conclusions you want to draw. If you want to give an overview of the actual SW situation - then using Cinebench is valid as one of many performance sampling points.
However if you want to reason about the microarchitecture, like if you want to conclude which architecture is more efficient - then such tests should be strictly excluded. Thats the whole point of the Spec CPU benchmark, it is intended to be used to compare different architectures.

 

[Xiao_Xi]

The geomean SPEC2017 scores for 7950X (full TDP) are 92.5 and 100.9 for int and fp tests, respectively. For M1 Max (also from anandtech benchmarks) its 53.3 and 81.07.

Out of curiosity, where did you find the geomean? I can't find them in the Anandtech articles.

Besides Phoronix, has anyone else benchmarked Zen 4 on Linux?

 

[leman]

Out of curiosity, where did you find the geomean? I can't find them in the Anandtech articles.

Calculated it myself.

 

[leman]

By the way, it's up to Apple to fix Embree's performance on their hardware and make sure the next version of Cinebench includes an ARM-optimized version.

They did submit a low-effort patch to embree, but it will probably take another year for Maxon to update Cinebench. But on a more serious note, why would Apple care much about embree's performance? They are investing into GPU-driven rendering and their work on Blender reflects it.

There are two or three users on these forums who judge performance by Cinebench and chess benchmarks, but they wouldn't buy Apple Silicon anyway. Rest of the world simply doesn't care. For me as a tech enthusiast the relevant question was why does Apple perform so much worse on CB23 than it does on virtually any other benchmark and now that I am fairly certain that I know the answer I don't find it interesting anymore.

 

[mi7chy]

No penalty for CPU benchmarks.

Better to wait for native Linux results since some tests have significant penalty under WSL2.

 

[leman]

(delete)

 

[mi7chy]

You go from not knowing what WSL is just a few posts back. We'll wait for more credible source.

 

[Gerdi]

Better to wait for native Linux results since some tests have significant penalty under WSL2.

View attachment 2081631

Dear, you should work on your reading comprehension. I did explicitly restricted my statement to CPU benchmarks, which not includes benchmarks testing I/O performance like an HTTP request benchmark.
Apparently you fail to understand the implications of running a system under virtualization or kernel emulation.

 

[pshufd]

"Fast"

A few benchmarks showed very little improvement.

No real analysis on power consumption or detailed component costs.

 

[diamond.g]

"Fast"

A few benchmarks showed very little improvement.

No real analysis on power consumption or detailed component costs.

HUB usually does motherboard reviews separately from the CPU reviews, so aside from a potential comment about how everything is expensive I wouldn't expect such a deep dive from them just yet.

Note: Gamers Nexus basically mentioned the high costs of the currently released X670E boards, which are more painful for the 7600 line than the 7950 (because the latter chip costs so much to begin with).

 

[mi7chy]

Dear, you should work on your reading comprehension. I did explicitly restricted my statement to CPU benchmarks, which not includes benchmarks testing I/O performance like an HTTP request benchmark.

Do you know what a localhost test server and client is? Here's another that's easier to understand. Or, do you want to test just CPU without system memory?

 

[Gerdi]

Do you know what a localhost test server and client is? Here's another that's easier to understand. Or, do you want to test just CPU without system memory?

Of course a CPU benchmark is using system memory - who would have guessed this - but NOT the network stack or other kernel services. It is really not that hard to grasp.

In summary the Spec CPU benchmark is still not affected by WSL (with a negligible impact coming from stage 2 translation) - despite your perceived counter-examples, which have nothing to do with the original question. That holds in particular, when using Java benchmarks (like Renaissance), where the applications has no control at all, what kernel services the JDK potentially requests.
A proper CPU benchmark like Spec does not have such dependencies.

 

[leman]

@mi7chy trying to talk to you is really frustrating. To show that WSL has a performance overhead you have to compare it to the equivalent test running under Windows directly. Instead you are throwing some obscure Java benchmarks at us that compare WSL Windows and standalone Linux. Not to mention that it's entirely unclear what these benchmarks are doing and how much the OS layer is involved.

 

[Xiao_Xi]

To show that WSL has a performance overhead you have to compare it to the equivalent test running under Windows directly.

Have you found a good benchmark to support that WSL has no overhead? Almost all benchmarks compare WSL with native Linux, rather than WSL with Windows. The only comparison between WSL and Windows that I have found suggests that Blender seemed to render faster on Windows than on WSL2.

Using CUDA for 3D Rendering on Windows vs WSL

Today, I am going to give you the premise of why I would go with Windows over WSL2 when it comes to 3...

dev.to

 

[leman]

Have you found a good benchmark to support that WSL has no overhead?

No, and I have no interest in looking. I fully admit that my statement that there is no overhead is only based on my experience and knowledge of computer systems. It is entirely possible that I am wrong and I will gladly accept that if someone provides evidence.

The only comparison between WSL and Windows that I have found suggests that Blender seemed to render faster on Windows than on WSL2.

Using CUDA for 3D Rendering on Windows vs WSL

Today, I am going to give you the premise of why I would go with Windows over WSL2 when it comes to 3...

dev.to

That's not a CPU-only test. It involves interfacing with the GPU which adds non-trivial complexity.

 

[mi7chy]

Blender does support CPU only along with GPU only and CPU+GPU. Here's 7950x on par with $2K msrp 32-core Threadripper 3970x.

https://www.profesionalreview.com/2022/09/26/amd-ryzen-9-7950x-review/

 

[Gerdi]

Have you found a good benchmark to support that WSL has no overhead? Almost all benchmarks compare WSL with native Linux, rather than WSL with Windows. The only comparison between WSL and Windows that I have found suggests that Blender seemed to render faster on Windows than on WSL2.

Using CUDA for 3D Rendering on Windows vs WSL

Today, I am going to give you the premise of why I would go with Windows over WSL2 when it comes to 3...

dev.to

You can just use Geekbench. Both under WSL and native Linux it will show about the same performance.

 

[mi7chy]

Geekbench is garbage though compared to real workloads.

Fortunately, Phoronix already did the comparison so I don't have to.

 

[Xiao_Xi]

It's crazy that the fight for the gaming crown has AMD and Intel pushing their CPUs past their sweet spot.

Intel Launches Raptor Lake at AMD: 24-core i9-13900K Arrives Oct 20 for $589

Intel fires back at AMD's Ryzen 7000

www.tomshardware.com

 

[pshufd]

It's crazy that the fight for the gaming crown has AMD and Intel pushing their CPUs past their sweet spot.

View attachment 2081916

Intel Launches Raptor Lake at AMD: 24-core i9-13900K Arrives Oct 20 for $589

Intel fires back at AMD's Ryzen 7000

www.tomshardware.com

That price point was interesting. I think that the 7950 is $600 and Intel is coming out just under that. I think that these are price cuts from when the previous generation models were introduced.

 

[kvic]

It's crazy that the fight for the gaming crown has AMD and Intel pushing their CPUs past their sweet spot.

I'm very happy to be alive and seeing Intel being an underdog for the 3rd time in the past 20 years:

Intel 13th gen is a laughing stock. Isn't it?