12/1/2020

Testing Methodology

• What we are monitoring
• How we are monitoring:
  • Power data
  • Temp sensors

Specs

• Computer specification
• OS
• Drivers
• Torch version

Anomolies

Suspicons

• Are things actually deterministic? There is a random number seed, but maybe something is still stocastic causing such variances in results. Would expect maybe 10%, but 2-3X is not expceted
• Is CUDA working as expected?
• Is there a background program
  • Checked top and htop, nothing other than standard practices
• Python version
• Pytorch

Ruled out

• Temps
• Power should be fine, no obvious issues, well below PSU specs
• Components themself → working fine in other programs