Cinebench 2024

[Werewolf]

I realise this is one step removed from chess, but it does provide useful data cross-platform and for Lc0 users.

Cinebench 2024 has been released. The GPU test has been added and is helpfully based on a similar test to the CPU one. It does require >8GB RAM on the card (or shared in Apple's case) which is why some systems below can't run it.

Apple MacBook Air M2
CPU = 382
GPU = 1582

Apple MacStudio M2 Ultra
CPU = 1922
GPU = 8658

i7-1260P
CPU = 418

i7-12700
CPU = 1042

Ryzen 3950X with 2080 Ti card
CPU = 1342
GPU = 8880

Threadripper 3990X with 4090 Card
CPU = 3384
GPU = 33139


Comments: Apple's ability to share memory helps them here, and their performance per watt (and size) is impressive. However, they cannot compete with fast desktop CPUs even from a few years ago and workstation CPUs are in another league. Strong Nvidia cards are, of course, much faster.

I think this provides a balanced picture of Apple's strength and weaknesses in an area considered to be their forte: graphics.

[dangi12012]

There are reviews around chesscode directly. They are ones of the few that test exactly that here:
https://www.tomshardware.com/reviews/am ... 287-5.html

While I do clearly see a correlation - for instance the impact of 3d-vcache is best observed directly on the type of code you aim for optimizing.

Also gpus are generally a huge untapped potential for computerchess as they rip through unconditional integer arithmetic like nobodys business - but you would need to run large parts of the engine on the device as well and then it becomes complicated with thread divergence and whatnot.
I am still working on testing at least the inference performance of NNUE when - lets say around 1024 accumulators exist in gpu memory to be evaluated.

[CornfedForever]

There are reviews around chesscode directly. They are ones of the few that test exactly that here:
https://www.tomshardware.com/reviews/am ... 287-5.html

While I do clearly see a correlation - for instance the impact of 3d-vcache is best observed directly on the type of code you aim for optimizing.

Also gpus are generally a huge untapped potential for computerchess as they rip through unconditional integer arithmetic like nobodys business - but you would need to run large parts of the engine on the device as well and then it becomes complicated with thread divergence and whatnot.
I am still working on testing at least the inference performance of NNUE when - lets say around 1024 accumulators exist in gpu memory to be evaluated.

Nice almost 3 yr old review of a chip.