Deployment & Scaling · Advanced
Quantisation and Inference Economics: FP8, FP4 and Cost per Million Tokens
Lower precision is the cheapest large win in GPU inference — but only if you measure quality alongside speed. Learn the formats NVIDIA hardware accelerates, the calibration workflow, and how to turn benchmarks into a cost-per-million-tokens figure a CFO can read.
- 1.6 hours lesson
- advanced level
- NVIDIA AI Track
- Narrated presentation
What you’ll learn
- FP16/BF16 baseline; INT8; FP8 on Hopper+; FP4-class on Blackwell; weight-only INT4
- Decode is bandwidth-bound, so fewer bytes per weight means faster generation
- Layer sensitivity and mixed-precision quantisation guided by calibration data
- A fixed evaluation set measured before and after — quality delta is mandatory, not optional
- Cost per million tokens as the executive-facing metric
- Utilisation as the hidden driver of cost per token
Inside this lesson
A narrated, slide-by-slide presentation with 1 prediction exercise, 1 knowledge check, 3 videos. You answer as you go, and the platform checks your work.