As @leonnowsden helpfully mentioned, make sure you’re on the latest version of Ubuntu and make sure that you’re plugged in when you’re performing long-running training jobs (otherwise you’ll run out of battery).
I have a brand new Tensorbook shipped with Ubuntu 18.04. It works fine when plugged in. But under battery power, it shuts down almost immediately after starting a GPU job. It works fine under battery power when the GPU is not in use. This issue is not fixed by fully charging the battery.
I’m guessing the batter cannot deliver the needed current at full voltage when the GPU is running. Anyone else have that? Any fixes known?
Hey Jimmy, one thing you can do is disable NVIDIA’s GPU for your display. This will reduce your load and may help with running your TensorBook with it not plugged in. Try running these commands @JimmyS:
Thanks for the suggestion. After that change, the GPU now does run on battery power without crashing the system.
But under batter power it’s running at about 1/3 the speed.
(This is measured in training samples processed per second. I’m running NetTrain on Wolfram Mathematica with GPU target). This effect is reversible and repeatable. I can unplug and plug in the power jack multiple times, and it will slow down and speed up accordingly each time.
This change also disables the NVIDIA X-server settings utility, which was handy for monitoring GPU use.