Apple M2 vs. AMD Ryzen 7 PRO 6850U Performance In Nearly 200 Benchmarks

[phoronix]

Phoronix: Apple M2 vs. AMD Ryzen 7 PRO 6850U Performance In Nearly 200 Benchmarks

Last week I published initial Apple M2 vs. AMD Rembrandt vs. Intel Alder Lake Linux benchmarks using Asahi Linux and Arch Linux across the board. For ending out this week, here is a follow-up article looking more closely at the Apple M2 in the MacBook Air against the AMD Ryzen 7 PRO 6850U "Rembrandt" within the Lenovo ThinkPad X13 Gen3. This time around are also results from performance tweaks to each laptop for the CPUFreq governor and platform profile.

https://www.phoronix.com/review/apple-m2-amd-ryzen

[timofonic]

I don't see M2 so magical as people reported in the past, but maybe Linux port needs more optimizations and better hardware support.

[fahrenheit]

M1 and M2 are only magical in perf/watt and that mostly if you compare with >25W cpus (like Intel which does beat the M2 in performance at the cost of power). There are indeed some situations were there are very large gains (see Techspot's review).

On the GPU side, the 680M is more or less comparable (which is actually a plus point for the M1/2). For gaming except for the very few games that exist on MacOS (and excluding iOS ports) the M2 prowess is moot because you can't actually play anything :P

For me a 6800U/6850U notebook would be perfect (better multithread, on par GPU), but the Mac does have battery life advantages (which for me don't matter much) and for better or worse a more coherent software ecosystem (yay for vertical integration).

[tchiwam]

Will be interesting to check the same benchmarks at 6 months, 1 year and 2 years from now

[Anux]

There are some workloads where the M2 loses hard like video encoding and liquid DSP but those are also workloads that could be done on one of its special hardware units. Which leads me to believe there was a bigger planning behind it while designing the CPU.

The real question is, will those special hardware units ever be supportet under linux?

[Indomitable]

I don't find it fair to test performance on M2 using Linux which is not optimized and GPU is not working. The better test would be testing M2 (Mac OS) again x64 (Linux).

[scottishduck]

I don't find it fair to test performance on M2 using Linux which is not optimized and GPU is not working. The better test would be testing M2 (Mac OS) again x64 (Linux).

Correct

[pkese]

I don't see M2 so magical as people reported in the past, but maybe Linux port needs more optimizations and better hardware support.

They seem pretty even, especially considering that AMD has 8 performance cores with hyper-threading and strong 256-bit AVX2 SIMD unit, whereas M2 has 4P+4E cores without HT and with much weaker 128-bit NEON SIMD.

It appears that AMD has already cashed in most of performance boosting features whereas Apple has more room to grow, especially in the SIMD and HT department.
Next year's architectures are likely to bring an advanced vector SIMD unit (SVE2) for Apple and some form ofAVX512 for AMD.

In the long run, I'd reckon that ARM ISA is going to deliver some 5% higher performance than X64 due to simpler instruction fetch and decode (less work to do, less power intensive, 1 stage shorter pipeline, 1 cycle less latency for mis-predicted branches, etc.) and less strict memory model.

[Anux]

What's not fair about that? It might not show you what you want to see but the article clearly states that it compares Linux. And Michael also doesn't hide the fact that Asahi is still a work in progress on Apple.
In fact every comparison is fair if no cheating was involved, you just have to keep in mind what you compare when making conclusions.

[GruenSein]

They seem pretty even, especially considering that AMD has 8 performance cores with hyper-threading and strong 256-bit AVX2 SIMD unit, whereas M2 has 4P+4E cores without HT and with much weaker 128-bit NEON SIMD.

It appears that AMD has already cashed in most of performance boosting features whereas Apple has more room to grow, especially in the SIMD and HT department.
Next year's architectures are likely to bring an advanced vector SIMD unit (SVE2) for Apple and some form ofAVX512 for AMD.

In the long run, I'd reckon that ARM ISA is going to deliver some 5% higher performance than X64 due to simpler instruction fetch and decode (less work to do, less power intensive, 1 stage shorter pipeline, 1 cycle less latency for mis-predicted branches, etc.) and less strict memory model.

I am also wondering about the role of SIMD units. Particularly in tests with huge performance differentials, there might be hand optimized AVX code paths but none for NEON since AVX is a lot more widely available. Simply compiling for ARM will not yield optimal usage of hardware resources in many cases.