The issue is it’s cpu compute which is underutilized in gpu clusters anyway, so practically it’s not really 1/1000.

Totally, edited my comment to specify "based on the assumption above." The main takeaway I was going for was 0.1% is not a small number in this context