AI-Float™ - Mixed Precision Arithmetic for AI: A Hardware Perspective

• 1. Introduction
• 2. Floating-point number representations
• 3. The IPU AI-Float^TM format
• 4. Lower precision for higher power efficiency
• 5. Mixed-precision arithmetic
• 6. Mixed-precision training
• 7. Deterministic versus stochastic rounding
• 8. Loss scaling
• 9. Summary
• 10. Trademarks & copyright