>"FP8, which has been used as the standard until now, was only supported by the RTX 40/50 series. In contrast,

INT8 ConvRot

is attracting attention because it can benefit from hardware support even on the RTX 20/30 series."

Related:

https://blog.comfy.org/p/minimax-h3-day-0-support-in-comfyui

Alternative title: "Our invention makes INT8 better on older consumer cards without FP8 support"

One thing is not obvious to me is how ConvRot can be applicable beyond diffusion models. Especially for LLM decoding, as each ConvRot would be more expensive for a given decoding vector, and it is required now, so you cannot easily get the benefit for prefill only, while maintaining the same decoding performance.