Apple M1 CPU & GPU speed is very disappointing

[Magnum]

There is a new very difficult discussion:
https://forums.macrumors.com/threads/ap ... 62/page-18
Start with page 18 today.
It could help the Stockfish and LC0 and Ceres developers a lot to improve these engines.

[Ras]

It could help the Stockfish and LC0 and Ceres developers a lot to improve these engines.

There's nothing even remotely helpful on Macrumors, as usually. Face it - Apple builds proprietary silicon that doesn't work well with chess engines in the first place, then slaps proprietary APIs around it that nobody cares about, and expects developers to buy these overpriced Apple toys instead of cross-compiling because Apple thinks that ripping of also developers and not just users is more profitable. In a word, don't expect improvement.

[Sopel]

[jdart]

That page is talking about Lc0, which is dependent on GPU speed. The M1 chips have an embedded GPU, which is not particularly powerful compared to dedicated external GPUs. It is not even clear Lc0 is compiled to use the native GPU, but even it was, you need big hardware to run Lc0 efficiently, beyond what you are going to find in a laptop.

[Raphexon]

Can you tell us how to optimize SF specifically for the M1?

Even if you can use the neural engine, it will likely incur a huge latency hit. So useless for SF.
NEON is still shit so that won't be faster either.
Maybe movegen optimization? ARM may like mailboxes more than bitboards, there's your 2% speed increase.

If you modify SF's movegen you can always do a pull request on the official stockfish github page.
Please add the results of a SPRT test with your M1 when you post it. Thanks!

[Ras]

Even if you can use the neural engine, it will likely incur a huge latency hit. So useless for SF.

Plus that there's no clear public documentation on how to even use it. Apple somehow thinks that keeping their hardware secret even from programmers will somehow lead to useful software. Don't ask for any sort of reason in that because that's Apple - "think different".

[wickedpotus]

That page is talking about Lc0, which is dependent on GPU speed. The M1 chips have an embedded GPU, which is not particularly powerful compared to dedicated external GPUs. It is not even clear Lc0 is compiled to use the native GPU, but even it was, you need big hardware to run Lc0 efficiently, beyond what you are going to find in a laptop.

Not a true statement.. you can get many Laptops with powerful RTX 30xx GPU that will run LC0 very efficiently. And furthermore most of these laptops will not cost significantly more than an Apple laptop only equiped with an on-chip M1 GPU.

[Milos]

That page is talking about Lc0, which is dependent on GPU speed. The M1 chips have an embedded GPU, which is not particularly powerful compared to dedicated external GPUs. It is not even clear Lc0 is compiled to use the native GPU, but even it was, you need big hardware to run Lc0 efficiently, beyond what you are going to find in a laptop.

Not a true statement.. you can get many Laptops with powerful RTX 30xx GPU that will run LC0 very efficiently. And furthermore most of these laptops will not cost significantly more than an Apple laptop only equiped with an on-chip M1 GPU.

He probably meant "what you are going to find in Apple laptop" .

[wickedpotus]

That page is talking about Lc0, which is dependent on GPU speed. The M1 chips have an embedded GPU, which is not particularly powerful compared to dedicated external GPUs. It is not even clear Lc0 is compiled to use the native GPU, but even it was, you need big hardware to run Lc0 efficiently, beyond what you are going to find in a laptop.

Not a true statement.. you can get many Laptops with powerful RTX 30xx GPU that will run LC0 very efficiently. And furthermore most of these laptops will not cost significantly more than an Apple laptop only equiped with an on-chip M1 GPU.

He probably meant "what you are going to find in Apple laptop" .

[Hai]

Can you tell us how to optimize SF specifically for the M1?

Even if you can use the neural engine, it will likely incur a huge latency hit. So useless for SF.
NEON is still shit so that won't be faster either.
Maybe movegen optimization? ARM may like mailboxes more than bitboards, there's your 2% speed increase.

If you modify SF's movegen you can always do a pull request on the official stockfish github page.
Please add the results of a SPRT test with your M1 when you post it. Thanks!

Yes NEON is shit. Only +18% speed improvement with one patch. https://abrok.eu/stockfish/
We deserve at least +180% with one patch.
I know that we have around thousands of speed patches for intel or amd cpus but that’s another story.