M1 Apple Silicon for Chess?

[Ckappe]

so - how are they exaggerated? I saw what I saw - my engine runs 1.5+ times faster on M1 (single thread) than on my 2700X and yes it's a desktop CPU and yes I'm talking 15W M1 vs 100W desktop. that's a fact - that M1 is 5nm and my 2700x is 12nm is not my concern.

You don't care about the generation of HW and nm-size.. but you care enough only to compare single-core performance only?.. the M1 a) costs you more than an old 2018 2700x and b) the 2700x will still beat it for chess-engine benchmarks if you choose to use the full-chip. Per core benches are really moot benchmarks for most engines today that happily does 256 threads if need be.

The only modern chess engine that would be helped by stronger per-core benches are LC0, Allie, etc. which are more GPU than CPU bound. But then the discussion will tilt to having the M1s built-in RAM+GPU compared to a comparable laptop CPU that supports discrete GPU and memory subsystems.

More comparable to M1 in CPU generation, price, and mobility is for example AMD Ryzen 7 PRO 5850U. It's similarly priced, 15W, and beats the M1 hands down in ANY relevant chess-engine-benchmark.

[Ras]

More comparable to M1 in CPU generation, price, and mobility is for example AMD Ryzen 7 PRO 5850U.

The one thing where it isn't comparable is availability. And with Ryzen 5000 laptops, you have to watch closely because 5xy0 with even x is Zen 3, but with odd x, it's older Zen 2 chips (shady AMD marketing). With the "PRO" variants - unless you're a big company that needs remote administration while bypassing the OS, that sounds more like a long term liability than an asset. I wouldn't want that in my hardware. Wait until 5800U becomes available, and then pay attention that you get 3200 MHz RAM if socketed, or 4600 LPDDR if soldered.

[mar]

You don't care about the generation of HW and nm-size.. but you care enough only to compare single-core performance only?.. the M1 a) costs you more than an old 2018 2700x and b) the 2700x will still beat it for chess-engine benchmarks if you choose to use the full-chip. Per core benches are really moot benchmarks for most engines today that happily does 256 threads if need be.

of course I can only compare the hardware I own and I compare my engine, because (surprise!) I don't care about SF or LC0.
why? obviously because my engine is way more relevant to me (and only me)

SF probably benefits quite a bit from AVX2 but arm64 only has 128-bit vector registers.
LC0 needs a badass GPU, obviously and my desktop 1080GTX smokes the M1 GPU easily, which is not surprising

I will probably compare the SMP results for my engine out of curiosity, both 8T and 16T vs 8T on M1 to see if my 2700X really wins (it might)

last but not least - I don't trust rigged benchmarks so I prefer to conduct my own.

why I bought M1 is none of your concern but here you go:
- I can't install new xcode on my old mbp, because it requires big sur and I can't install it on my old mbp, because apple
what this implies? my old mbp has no support for Metal, so if I ever had to support it, I couldn't (again because apple)
- I can't build mt apps anymore for recent ios, because apple. I have some other apps I need to maintain in the future
- I cannot build a single x64+arm64 binary on my old mbp
- I wanted arm64 CPU, because it's not x86 (for reasons a mere user probably wouldn't understand)
- M1 absolutely smokes my old 2009 mbp in compile performance, but that was some very old dual-core intel mobile chip anyway

[mar]

SMP results for my engine

Code: Select all

19'725'019 nps, 2700X 16 threads
13'884'314 nps, 2700X 8 threads
17'497'848 nps, M1 8 threads

perf diff (left is faster):
desktop 16T vs desktop 8T: 1.42x (pretty good for hyperthreading)
desktop 16T vs M1: 1.13x
M1 vs desktop 8T: 1.26x

so even the MT performance is not that bad on M1, considering the slow cores.
2700X at full load wins

[Ckappe]

You don't care about the generation of HW and nm-size.. but you care enough only to compare single-core performance only?.. the M1 a) costs you more than an old 2018 2700x and b) the 2700x will still beat it for chess-engine benchmarks if you choose to use the full-chip. Per core benches are really moot benchmarks for most engines today that happily does 256 threads if need be.

of course I can only compare the hardware I own and I compare my engine, because (surprise!) I don't care about SF or LC0.
why? obviously because my engine is way more relevant to me (and only me)

SF probably benefits quite a bit from AVX2 but arm64 only has 128-bit vector registers.
LC0 needs a badass GPU, obviously and my desktop 1080GTX smokes the M1 GPU easily, which is not surprising

I will probably compare the SMP results for my engine out of curiosity, both 8T and 16T vs 8T on M1 to see if my 2700X really wins (it might)

last but not least - I don't trust rigged benchmarks so I prefer to conduct my own.

why I bought M1 is none of your concern but here you go:
- I can't install new xcode on my old mbp, because it requires big sur and I can't install it on my old mbp, because apple
what this implies? my old mbp has no support for Metal, so if I ever had to support it, I couldn't (again because apple)
- I can't build mt apps anymore for recent ios, because apple. I have some other apps I need to maintain in the future
- I cannot build a single x64+arm64 binary on my old mbp
- I wanted arm64 CPU, because it's not x86 (for reasons a mere user probably wouldn't understand)
- M1 absolutely smokes my old 2009 mbp in compile performance, but that was some very old dual-core intel mobile chip anyway

I understand that "if" you want to build iOS apps, it for sure makes sense with a Mac, especially as this is a very locked-in eco-system. Older MBPs are not better in terms of price/performance.

That said this thread is about the M1s merits for Chess, and it makes the most sense to compare it with current alternatives cheap, similar, or slightly higher priced mobile system options. Comparing it only within the restricted, narrow constraints of Apple-branded products, makes little sense, for me as a chess fan.

Even according to your own benchmarks on your engine the inexpensive 2700x beats a spanking new M1.. So what is really to be impressed with?.. Yes.. Its a more power-consuming Desktop CPU... But its older and the more relevant comparison today is for example with a 5xxxU 15W AMD mobile CPU...

The point made about the GPU is that the M1 is what it is, it's a system on a chip, every GPU-bound application will be restricted to how "slow" the M1 is in this regard. Other systems costing similar to the M1 even offer both onboard-GPU and discrete "badass" GPUs more powerful than your current desktop in a slim inexpensive laptop today. I think this is relevant for most chess fans as there are both GPU-bound and CPU-bound popular chess engines these days. And if you are a developer that focuses on more than just iOS, having more options and AI and graphics capabilities is certainly not a drawback these days.

I completely agree with you on benchmarks. Geekbench-scores etc. have proven to be pretty useless when trying to compare different architectures (which was their original goal btw). And media, marketing abuses the benchmarks. I think many of these media/marketing-driven "bad" benches for M1 have contributes to it being hyped beyond what it delivers in terms of performance.

[JohnWoe]

I don't see any cockblocking that Apple users aren't allowed to enjoy their devices.

lol ARM is perfect and x86 is crap.
For me the deal breaker is the lack of USB-A ports.
That's why ThinkPad!
Despite the stupid Windows tax.

Never owned any Apple device, so I can't comment their OS.

[Ckappe]

I don't see any cockblocking that Apple users aren't allowed to enjoy their devices.

lol ARM is perfect and x86 is crap.
For me the deal breaker is the lack of USB-A ports.
That's why ThinkPad!
Despite the stupid Windows tax.

Never owned any Apple device, so I can't comment their OS.

With current "apple-tax" The 17inch 144hz Legion 5 laptop with 5800H 8/16 cores, 16GB RAM, 1TB ssd, RTX3070, actually costs LESS!! on most markets than a Macbook Air M1 with 16GB RAM and 512GB ssd.

[AlexChess]

I don't see any cockblocking that Apple users aren't allowed to enjoy their devices.

lol ARM is perfect and x86 is crap.
For me the deal breaker is the lack of USB-A ports.
That's why ThinkPad!
Despite the stupid Windows tax.

Never owned any Apple device, so I can't comment their OS.

With current "apple-tax" The 17inch 144hz Legion 5 laptop with 5800H 8/16 cores, 16GB RAM, 1TB ssd, RTX3070, actually costs LESS!! on most markets than a Macbook Air M1 with 16GB RAM and 512GB ssd.

I've answered to all your M1 hardware related comments here: http://talkchess.com/forum3/viewtopic.p ... 62#p889462

[mar]

Even according to your own benchmarks on your engine the inexpensive 2700x beats a spanking new M1.. So what is really to be impressed with?.. Yes.. Its a more power-consuming Desktop CPU... But its older and the more relevant comparison today is for example with a 5xxxU 15W AMD mobile CPU...

are you serious? my tests show that 8T vs 8T M1 is still significantly better (to my surprise), and that's a mobile chip running on battery - so yes, it's actually very impressive
note that I never run self tests with hyperthreading, so it's actually relevant to me (not that I plan to waste my M1 on chess, but I might)

I understand you probably have some personal grudge against M1 for some reason, but so far I'm very happy with it and you nor anyone else can spoil that. and that's just the first generation.
not to mention that competition is always good for us as it pushes others to do better

EDIT: the 2700X needs 13T to actually start beating M1 8T

[Ckappe]

not that I plan to waste my M1 on chess, but I might

That sums it up, I suppose. The thread is "M1 Apple Silicon for chess".

I have no personal grudge against any HW. I just think it is borderline fanboyism to compare HW and then don't optimize benches based on the HW used on more than one end. Of course, HT should be used on CPUs using it, everything else is limiting one CPU and optimizing for the other.

I also have no gripes about you being "impressed" with whatever you like. I am just putting forward the notion that the price/performance equation of the M1 hardware pretty much sucks for Chess as of today (compared to current generation competing products). A similar priced modern AMD laptop will BEAT both your 2700x desktop and your M1 laptop both on GPU-chess and CPU-chess, regardless of bench you throw at it.

Why not as well run your bench with both CPUs at 16T. if you don't care about optimizing benches for each CPUs