M1 Max Thermal Throttling

[yadamonk]

Hi @tlkh

Thank you so much for creating and running those benchmarks. I would be interested if thermal throttling affects training speed (ms/step) after a few minutes.

Could you share your training logs when running e.g. bm_rn50.py for benchmark_epochs = 20?

[yadamonk]

Alternatively an asitop screenshot taken during a longer benchmark run could also tell us if GPU power draw declines over time.

[tlkh]

Will do, however, I can say for now that

1. the fans barely even spin up
2. there's no easy way to monitor temperature yet

[yadamonk]

Thank you for that information. I suspect the fans will ramp up more after 5-10 minutes.

[tlkh]

Ran for about 10 mins. You're right, I did hear the fan this time (but it is very faint!)

Don't observe any throttling, but the perf fluctuates a bit probably due to OS and background tasks etc.

[yadamonk]

Thank you so much for sharing. These are really impressive results.

[yadamonk]

Just out of curiosity, what happens when you run the M1 Max in low power mode. I'm especially interested in the power consumption of the GPU and the score you get for bm_rn50.py.

[yadamonk]

The background for my question is that I'm wondering if low power mode might be an option for longer training runs e.g. over night.

[tlkh]

Sorry, recently I bit more occupied with other stuff, I plan to circle back and test it later this week

[yadamonk]

Today I finally got my M1 Max and I'm happy to report that for the resnet50 benchmark low power mode achieves 61.4% of the performance you reported with an average GPU power consumption of only 15 Watts (total package power is around 30).

Needless to say the machine stays cool as a cucumber even when training for longer which is really important to me as it's my primary computer.