M1 Max Ethereum Mining Test

[jmurphyau]

With 400 GB/s of memory, and with 8,192 bytes of memory accessed for every hash, I would expect to see a hash rate of 48m - so there seems to be room to push this I think.. Given I was able to get 100m+ using cache, and the only change was memory, I think it's just about how the memory is accessed

I compared the memory bandwidth on my RTX 3090 (936gb/s) and used the same calculation (8,192 bytes per hash) and the result is 114m - which is close to what I see on that card (it's reporting 107m)

 

[diamond.g]

With 400 GB/s of memory, and with 8,192 bytes of memory accessed for every hash, I would expect to see a hash rate of 48m - so there seems to be room to push this I think.. Given I was able to get 100m+ using cache, and the only change was memory, I think it's just about how the memory is accessed

I compared the memory bandwidth on my RTX 3090 (936gb/s) and used the same calculation (8,192 bytes per hash) and the result is 114m - which is close to what I see on that card (it's reporting 107m)

Are you able to modify the way memory is accessed?

 

[Kedano]

Has there been any improvements to this @jmurphyau @ChainfireXDA ?

 

[diamond.g]

I'd be curious (cause clearly this is a Eth postmortem, lol) if M3's new architecture fixes whatever was causing the bottleneck in M1/M2 (assuming M2 didn't fix anything). @leman what do you think?