Apple M1 CPU & GPU speed is very disappointing

[Appletoni]

Benchmark: Stockfish (chess) speed

M1 CPU = 13000 kn/s

i7 3930k overclocked (from 2011 = 10 years old) = 13000 kn/s

Others = 40000 kn/s - 80000 kn/s

Desktop CPUs = 230000 kn/s and much stronger

 

[leman]

I have no idea what Stockfisch is or why it should be relevant for performance comparisons. I do know that M1 performs on par with enthusiast level desktop CPUs in industry standard benchmarks such as SPEC.

 

[Andropov]

Yes, M1 CPU seems to be way below (its) average on this particular benchmark. If your main use for your M1 Mac is to run chess engines, you would be better off buying a different computer.

 

[Fomalhaut]

Benchmark: Stockfish (chess) speed

M1 CPU = 13000 kn/s

i7 3930k overclocked (from 2011 = 10 years old) = 13000 kn/s

Others = 40000 kn/s - 80000 kn/s

Desktop CPUs = 230000 kn/s and much stronger

It would be interesting to know exactly how this benchmark works in order to understand why the M1 chip does poorly. It the benchmark running natively or under Rosetta 2?

In all other benchmarks that have been published (that I am aware of), the M1 performs exceptionally well compared to other CPUs with a similar TDP.

 

[CMMChris]

Well, thats the thing with different CPU architectures. The one might perform better than the other in very specific workloads. ARM has its weak spots just like x86 has. However, that does not mean that the CPU is slow. Overall real-life performance is what matters in the end.

 

[jeanlain]

Benchmark: Stockfish (chess) speed

M1 CPU = 13000 kn/s

i7 3930k overclocked (from 2011 = 10 years old) = 13000 kn/s

Others = 40000 kn/s - 80000 kn/s

Desktop CPUs = 230000 kn/s and much stronger

Source?
And why should we trust this tool to compare CPU speed? It's not a standard benchmark, it's a chess engine. We don't know if the code is platform agnostic, but it most likely isn't.
If you want to compare CPU speed, use industry-standard benchmark tools like SPEC. Geekbench is also more representative.

 

[Andropov]

Looks like Stockfish doesn't use the Neural Engine in the M1 (which is expected, since it's a multiplatform tool apparently). If you could find a chess engine that uses the Neural Engine (using CoreML) the results might be quite different.

 

[JouniS]

It's easy to find real-world examples where the M1 loses to Intel/AMD for predictable reasons.

Stockfish apparently caches previously evaluated positions in giant tables, and because the M1 only supports 16 GB memory, it may have to do redundant work that systems with more memory can avoid. Stockfish also uses plenty of SIMD and bit manipulation instructions, which may not always have good ARM equivalents. Finally, combinatorial algorithms are often very sequential in nature, and high instruction-level parallelism doesn't help much with such code.

 

[Gnattu]

Stockfish engine is optimized for x86's SIMD instructions extensively, and its algorithm cannot utilize M1's super-wide ALU effectively, this is why you are having "disappointing speed." If chess analysis is your main use case you should not buy a M1 Mac and you have to stick to x86 for now.

 

[Apple_Robert]

Benchmark: Stockfish (chess) speed

M1 CPU = 13000 kn/s

i7 3930k overclocked (from 2011 = 10 years old) = 13000 kn/s

Others = 40000 kn/s - 80000 kn/s

Desktop CPUs = 230000 kn/s and much stronger

Benchmarking is a waste of time. What matters, is real world day to day use.

 

[snorkelman]

Benchmarking is a waste of time. What matters, is real world day to day use.

I think benchmarking *is* some folks real world day to day use

 

[TiggrToo]

Benchmark: Stockfish (chess) speed

M1 CPU = 13000 kn/s

i7 3930k overclocked (from 2011 = 10 years old) = 13000 kn/s

Others = 40000 kn/s - 80000 kn/s

Desktop CPUs = 230000 kn/s and much stronger

Biig freaking whoop. It's a benchmark test running x64 under Rosetta 2.

In other words, a totally useless and pointless measurement of M1 chips and simply an indication of strengths and weakness of Rosetta 2.

 

[mtneer]

That is to be expected. Not every tool is best for every possible use condition.

 

[Apple_Robert]

I think benchmarking *is* some folks real world day to day use

When it comes to many on MR, I have no doubt that is the case. The sad thing is many of these same people will return a perfectly good M1 because of what the benchmark test showed. Crazy stuff.

 

[snorkelman]

When it comes to many on MR, I have no doubt that is the case. The sad thing is many of these same people will return a perfectly good M1 because of what the benchmark test showed. Crazy stuff.

I know utter madness:

"Look darling, it's just rendered our *entire* honeymoon video in 18 minutes! Isn't that fantastic? You must be delighted that I bought it for our anniversary"

'Oh really? Well just you put that honeymoon video aside and take a good long look at the Stockfish benchmark - only 13000 something or others. It's *all* ruined, I want a divorce!!'

 

[jdb8167]

Biig freaking whoop. It's a benchmark test running x64 under Rosetta 2.

In other words, a totally useless and pointless measurement of M1 chips and simply an indication of strengths and weakness of Rosetta 2.

It doesn’t have to be. It is available from Homebrew and is labeled as Apple Silicon. That doesn’t mean it is particularly optimized for Aarch64 though.

 

[mi7chy]

Not bad. My phone only does ~2,300k n/s with Stockfish engine configured for four threads. When I get back I'll try it on an AMD 5950x 16-core.

Comprehensive list of Stockfish benchmarks here with AMD Epyc and Threadripper dominating the top positions with 282,000k n/s for 2x Epyc 7763.

https://openbenchmarking.org/test/pts/stockfish

 

[leons]

After reading this, I now realize how poorly my Macbook M1 will perform with Chess. I am devastated and don't know what I will do to fix this situation. (What's Chess)?

 

[snorkelman]

(What's Chess)?

Dunno, think it's some game they're releasing to cash in on a popular benchmark

 

[matrix07]

After reading this, I now realize how poorly my Macbook M1 will perform with Chess. I am devastated and don't know what I will do to fix this situation. (What's Chess)?

I'm still not sure if the OP really serious. ?

 

[spiderman0616]

Yeah, I'm super disappointed in my M1 Mac that never gets more than room temperature, performs every bit as fast as (if not faster than) my iPad and iPhone, and only needs to be charged about once a week. What a garbage machine.

 

[darngooddesign]

Benchmark: Stockfish (chess) speed

M1 CPU = 13000 kn/s

i7 3930k overclocked (from 2011 = 10 years old) = 13000 kn/s

Others = 40000 kn/s - 80000 kn/s

Desktop CPUs = 230000 kn/s and much stronger

Thats super disappointing since ...

 

[BigMcGuire]

How are you benchmarking this? The kN/s varies wildly depending on position.

On my M1 MBP after I played e4 -

with the settings:

.

Mind you, the default setting was 1 thread and 128MB hash.

--------------------------------------------------------------------

My 2019 MBP 15' with an i7-9750H (6 cores, 12 threads) @ 2.60Ghz with the same preference settings (max threads) got 9232kN/s with the same position. Fans immediately turned on and started screaming. Pic below.

 

[mi7chy]

Funny how people selectively preach dorkbench when M1 does well but downplay a more relevant real world workload when M1 doesn't. Chess has been a relevant workload and benchmark going back to IBM Deep Blue to current DeepMind AlphaZero that topple all the grandmaster human players.

There was maybe one intelligent response trying to understand what the bottleneck is.

 

[xraydoc]

Funny how people selectively preach dorkbench when M1 does well but downplay a more relevant real world workload when M1 doesn't. Chess has been a relevant workload and benchmark going back to IBM Deep Blue to current DeepMind AlphaZero that topple all the grandmaster human players.

There was maybe one intelligent response trying to understand what the bottleneck is.

No, the issue is not downplaying unfavorable benchmarks, but the benchmarks must be fair and properly coded for the ARM/M1 platform. While I know nothing of Stockfish, it seems not to be coded well if the M1 is getting 1/10th the speed of comparable x86 processors, yet the M1 holds ground on virtually everything else known to be native ARM code.