This could be good for chess

[Cardoso]

https://www.youtube.com/watch?v=B635wcdr6-w

[hgm]

This appears to be the famous proprietry TPU hardware that was used by Google / Deep Mind to run AlphaZero, which now has gone commercial. It seems you can buy it for 1232 euro at Rutronik (German?), or $1195 at ShopBLT (but probably import duty on that in Europe?).

This should be much more powerful for running neural nets than any GPU.

[Joost Buijs]

This appears to be the famous proprietry TPU hardware that was used by Google / Deep Mind to run AlphaZero, which now has gone commercial. It seems you can buy it for 1232 euro at Rutronik (German?), or $1195 at ShopBLT (but probably import duty on that in Europe?).

This should be much more powerful for running neural nets than any GPU.

I've been looking at this some time ago. Compared to the TPU hardware that Google used for training and running Alpha Zero this is toy hardware.

When you take the version with 8 Coral Edge M2 TPU modules it theoretically does 64 (8 bit) TOPS.

https://dlcdnets.asus.com/pub/ASUS/mb/A ... c_1225.pdf

A modern GPU based on the NVidia Ampere architecture like the A100 theoretically does 624 (8 bit) TOPS, that is almost 10 times as fast.

[dangi12012]

This appears to be the famous proprietry TPU hardware that was used by Google / Deep Mind to run AlphaZero, which now has gone commercial. It seems you can buy it for 1232 euro at Rutronik (German?), or $1195 at ShopBLT (but probably import duty on that in Europe?).

This should be much more powerful for running neural nets than any GPU.

I've been looking at this some time ago. Compared to the TPU hardware that Google used for training and running Alpha Zero this is toy hardware.

When you take the version with 8 Coral Edge M2 TPU modules it theoretically does 64 (8 bit) TOPS.

https://dlcdnets.asus.com/pub/ASUS/mb/A ... c_1225.pdf

A modern GPU based on the NVidia Ampere architecture like the A100 theoretically does 624 (8 bit) TOPS, that is almost 10 times as fast.

Not even Theroretically. My RTX 3080 does 120Tflops end to end for fp16 caluclations and over 1.8 petaOPS for binary inpout. Thats about 2 Billion positions/second for a network size similar to NNUE WITHOUT incremental updates.

You heard it here first: If someone writes a full chess engine in CUDA with similar move ordering and pruning to SF - it will break all ELO records. Also Gpus are very very scalable since one rtx 3080 is already faster than a 5950x for integer and float math - and you can always use multiple gpus in a system.

Oh my oh my - we are still at the very beginning of computer chess. Maybe to compare different hardware the notion of ELO/Watt or Meganodes/s/watt is best for the long term future.

[Joost Buijs]

This appears to be the famous proprietry TPU hardware that was used by Google / Deep Mind to run AlphaZero, which now has gone commercial. It seems you can buy it for 1232 euro at Rutronik (German?), or $1195 at ShopBLT (but probably import duty on that in Europe?).

This should be much more powerful for running neural nets than any GPU.

I've been looking at this some time ago. Compared to the TPU hardware that Google used for training and running Alpha Zero this is toy hardware.

When you take the version with 8 Coral Edge M2 TPU modules it theoretically does 64 (8 bit) TOPS.

https://dlcdnets.asus.com/pub/ASUS/mb/A ... c_1225.pdf

A modern GPU based on the NVidia Ampere architecture like the A100 theoretically does 624 (8 bit) TOPS, that is almost 10 times as fast.

Not even Theroretically. My RTX 3080 does 120Tflops end to end for fp16 caluclations and over 1.8 petaOPS for binary inpout. Thats about 2 Billion positions/second for a network size similar to NNUE WITHOUT incremental updates.

You heard it here first: If someone writes a full chess engine in CUDA with similar move ordering and pruning to SF - it will break all ELO records. Also Gpus are very very scalable since one rtx 3080 is already faster than a 5950x for integer and float math - and you can always use multiple gpus in a system.

Oh my oh my - we are still at the very beginning of computer chess. Maybe to compare different hardware the notion of ELO/Watt or Meganodes/s/watt is best for the long term future.

The Asus accelerator needs 16 Coral Edge M2 modules to reach 64 TOPS and not 8 like I said before. I compared it with the NVidia A100 because it's specs are pretty well known.

NVidia claims that the A100 does 624 TOPS with 8 bit tensors (1248 TOPS with scarce matrix multiplication). Over here I use a RTX 3090 (turbo), this card has approx. 2/3 of the A100 performance, my guess is that it does around 400 TOPS which is still 6.25 times more powerful than the Asus accelerator.