[KongchangAI]
ConceptW8A8

int8 W8A8

A quantization scheme in which both weights and activations are quantized to 8-bit integers, requiring hardware support for int8 matrix multiplication.

Timeline (last 90 days)

Nov 6

On AMD RDNA3/RDNA4 GPUs with ROCm, compiled int8 W8A8 linear layers make Qwen-Image-2.1 int8 21% faster per step.

Unverified50%

All Facts (1)

Source Articles