M1 Apple Silicon for Chess?

[Ras]

The most interesting part will be to know if the new Apple computers will have options of an added on-board GPU-chipset (Nvidia or AMD)

That would make little sense for Apple. Their move is to integrate all of it, not buying from others, so I'd expect them to use an upscaled version of the on-chip graphics they have now. The other question is - for what? Gaming is likely out of question anyway.

Let's not forget AMD already has 5700G CPU released with stellar performance, extremely good power efficiency and with killer graphics on-chip.

But no availability until August, and it remains to be seen how the supply will look then. Probably not good because APUs are not that attractive for AMD to manufacture, compared to normal Ryzens. That's because the APU Ryzens are monolithic, not with chiplets, so you get fewer Ryzens per wafer.

Also, the 5700G graphics aren't better than in the 4750G, and these are marginally better than in the 3400G (4750G: DDR4-3200, 3400G: DDR4-2933). That's because all of them are bandwidth starved with DDR4. You get about GT 1030 performance, which isn't killer graphics. It's OK for office and multimedia, or maybe old games at low settings, but that's it.

[wickedpotus]

That would make little sense for Apple. Their move is to integrate all of it, not buying from others, so I'd expect them to use an upscaled version of the on-chip graphics they have now. The other question is - for what? Gaming is likely out of question anyway.

But no availability until August, and it remains to be seen how the supply will look then. Probably not good because APUs are not that attractive for AMD to manufacture, compared to normal Ryzens. That's because the APU Ryzens are monolithic, not with chiplets, so you get fewer Ryzens per wafer.

Also, the 5700G graphics aren't better than in the 4750G, and these are marginally better than in the 3400G (4750G: DDR4-3200, 3400G: DDR4-2933). That's because all of them are bandwidth starved with DDR4. You get about GT 1030 performance, which isn't killer graphics. It's OK for office and multimedia, or maybe old games at low settings, but that's it.

Some very good points there, Finding the right chipset for Chess-engines usually converges with the best gaming-computer options, in my opinion - the exception being some very strong server-hw-setups. Especially these days when the strongest engines are both GPU and CPU bound, for training and playing.

[wickedpotus]

Tomorrow at WDCC21 Apple will show new Silicon Mx devices 2-3 times faster, but still slims, silents and freshes. AMD like INTEL and Microsoft with Windows 11 are moving to ARM, that is the future

So after yesterday's lack of announcement, it seems Apple silicon may be stuck with the 4-core (4 fully powered ones) M1 they have today. And will not get any closer to the already available Ryzens at least before autumn.

[Magnum]

Tomorrow at WDCC21 Apple will show new Silicon Mx devices 2-3 times faster, but still slims, silents and freshes. AMD like INTEL and Microsoft with Windows 11 are moving to ARM, that is the future

So after yesterday's lack of announcement, it seems Apple silicon may be stuck with the 4-core (4 fully powered ones) M1 they have today. And will not get any closer to the already available Ryzens at least before autumn.

I have the Macbook Pro M1 and it’s great.
I will buy the 16 inch too.
But I’m waiting for a MacBook Pro 16 inch (or better 18 inch) and 16 CPU performance cores or more to use Stockfish.

To use LC0 I would need a MacBook Pro with 1032 or 2064 GPU cores.

See discussions here
https://github.com/official-stockfish/S ... ssues/3529
https://forums.macrumors.com/threads/ap ... 62/page-12
= It’s a problem of Apple M1 hardware, no it’s a problem of Stockfish code, no it’s…

[wickedpotus]

= It’s a problem of Apple M1 hardware, no it’s a problem of Stockfish code, no it’s…

Stockfish is c++ and a specially optimized version for ARM has been developed using for example NEON-instructions, just as Stockfish developers all around the world in general tries to optimize code and compile for most modern, popular architectures.

Stockfish is open source so anyone claiming the code could be optimized "better" for Apple silicon are free to prove their point!! Nothing stopping the people to put their check-outs and commits where their mouth is in this regard.

This brings back the "old" discussion of CISC vs RISC... Whats *best*.. a richer set of instructions that take more clock-cycles to perform but on the other hand can get more stuff done and be specialized for certain types of useful operations (and can be cleverly used by programmers and compilers), or a fewer set of simpler instructions that are very fast but may need to be combined heavily to accomplish what more complex instructions can do in one op with a more complex set of specialized op-codes.

Most chess-lovers however don't care about why the M1 performs poorly compared to cheaper alternatives. Most chess-engine-users just care what CPU gives you the most chess-power for your buck. And the M1 seems severely overhyped by Apple and mainstream media based on some very questionable benchmarks, that suit ARMs instruction-set and may not have been cleverly optimized to make the best use of the CPUs with richer instruction sets, like stockfish and some high-end game engines etc.

[acepoint_de]

Five updates for the m1: berserk-4.2.0, Ethereal-13.0.0, honey-13.2.1, mayhem-5.0, shashchess-17.1

https://acepoint.de/five-updates-for-the-m1/

Ciao

acepoint

[AlexChess]

Five updates for the m1: berserk-4.2.0, Ethereal-13.0.0, honey-13.2.1, mayhem-5.0, shashchess-17.1

https://acepoint.de/five-updates-for-the-m1/

Ciao

acepoint

Thank you my skilled chess friend! Finally I can test Ethereal 13.00 at full speed!!! (ShashChess 17.1 I already know that is very strong, maybe the new number 1)
https://db5pap001files.storage.live.com ... height=204

[Magnum]

Five updates for the m1: berserk-4.2.0, Ethereal-13.0.0, honey-13.2.1, mayhem-5.0, shashchess-17.1

https://acepoint.de/five-updates-for-the-m1/

Ciao

acepoint

Thank you my skilled chess friend! Finally I can test Ethereal 13.00 at full speed!!!

What do you mean with full speed? What was the problem before?

The M1 seems severely overhyped by Apple and mainstream media based on some very questionable benchmarks, that suit ARMs instruction-set and may not have been cleverly optimized to make the best use of the CPUs with richer instruction sets, like stockfish and some high-end game engines etc.
= Apple can improve some things.

But Stockfish needs improvements too, look what the Stockfish developers said:
-Stockfish does support NEON but that's about it, all the optimizations goes towards x86 SIMD such as AVX and BMI.
-If you want better support for its instruction set you'll have to code it yourself. And like with other ARM chips, it does simple instructions incredibly well but starts losing ground with more complex ones.
-The Stockfish community would welcome any developer to come in to help with coding for the M1.
-There's no optimization yet that uses the additional hardware available on M1 (e.g. the rumored matrix-multiplier or the neural engine). A way to use that could be by accessing Apple's accelerate framework (will be shown in today's lesson at WWDC). So there definitely is potential, lot's of unknowns, and for sure no easy task.
-We can agree that SF isn't optimized for M1 chips.
-If somebody comes up with better code, specific for M1, we'll happily take a look.
-very simple reason... nobody made a PR for M1. Simply do it..

= Stockfish could be up to 10 times faster

[wickedpotus]

= Stockfish could be up to 10 times faster

Very unlikely, and remember stockfish-code could also be optimized and tweaked for the already quicker Ryzens etc. as well. Assuming Stockfish is "less" optimized for an M1 compared to a threadripper is a false narrative. There are probably even more optimizations still there for CPUs with complex instruction sets over ARM CPUs with more limited set of OPs, and generally less optimizations avenues.

[AlexChess]

Five updates for the m1: berserk-4.2.0, Ethereal-13.0.0, honey-13.2.1, mayhem-5.0, shashchess-17.1

https://acepoint.de/five-updates-for-the-m1/

Ciao

acepoint

Thank you my skilled chess friend! Finally I can test Ethereal 13.00 at full speed!!!

What do you mean with full speed? What was the problem before?

The M1 seems severely overhyped by Apple and mainstream media based on some very questionable benchmarks, that suit ARMs instruction-set and may not have been cleverly optimized to make the best use of the CPUs with richer instruction sets, like stockfish and some high-end game engines etc.
= Apple can improve some things.

But Stockfish needs improvements too, look what the Stockfish developers said:
-Stockfish does support NEON but that's about it, all the optimizations goes towards x86 SIMD such as AVX and BMI.
-If you want better support for its instruction set you'll have to code it yourself. And like with other ARM chips, it does simple instructions incredibly well but starts losing ground with more complex ones.
-The Stockfish community would welcome any developer to come in to help with coding for the M1.
-There's no optimization yet that uses the additional hardware available on M1 (e.g. the rumored matrix-multiplier or the neural engine). A way to use that could be by accessing Apple's accelerate framework (will be shown in today's lesson at WWDC). So there definitely is potential, lot's of unknowns, and for sure no easy task.
-We can agree that SF isn't optimized for M1 chips.
-If somebody comes up with better code, specific for M1, we'll happily take a look.
-very simple reason... nobody made a PR for M1. Simply do it..

= Stockfish could be up to 10 times faster

Ethereal 13.00 and Pedone 3.1 are the only engines (tested 50+ until now) that doesn't run fine on my Windows 10 ARM64 VM under Parallels Desktop M1, that is compatible with x64 popcnt ssse3|4.2 (but not avx avx2 or bm2). I have discovered that they do not crash if I set only 1 CPU for them. Reported that to Andrew Grant, but he has answered that it doesn't support emulated systems. I can understand that, because it is tryng to get the most from his new engine using latest estensions. (Note: Ethereal 12.75 works perfecty on my system, also with Stockfish TEST NNUEs. I'm about to publish here a crosstable with a blitz match between 38 engines on that hardware) Acepoint kindly compiled the github version for Mac M1. So insted than only 1 CPU, now I can test it with up to 8 CPUs. Sadly now only the commercial version has a proprietary NNUE, but the challenge against all the other top engines will be hardware-fair using the free github version. I'll report the results here, too. And I bet that Ethereal 12.75 + MikeB's c4c3... NNUE will be stronger than Ethereal 13.00 standard and maybe very close to Ethereal 13.00 NNUE commercial that is 20 Elo point stronger than Igel 3.0.5 NNUE, according to most engine testers: http://talkchess.com/forum3/viewtopic.p ... 20#p895711