Stories tagged with #floating-point

Found 1 story

johndcook.com — FP4: the 16-value float powering billion-parameter models
1.4k 1.4k

FP4: the 16-value float powering billion-parameter models

FP4, a 4-bit float with just 16 representable numbers, is now the default inference format for billion-parameter models on NVIDIA Blackwell hardware.

machine-learning floating-point nvidia quantization ai-inference
johndcook.com