SafetyPage

Evaluators and safety organisations · Article · Epoch AI

Changing how we measure AI chip inference performance

Read at epoch.ai ↗Read at sourceAll rights reserved

Outline · 52 paragraphs

  1. The tradeoff between concurrency and output speed¶7
  2. How we calculate the IPI¶15
  3. Comparing capacity across output speeds¶27
  4. Combining the model comparisons¶32
  5. Interpreting the IPI¶36
  6. Acknowledgements¶42