Comment by boroboro4

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