INT8 ConvRot Method Proposed as Alternative to FP8 in Neural Networks

A technical article published on Note.com introduces a numerical method called INT8 ConvRot, which its author claims could replace FP8 in machine learning computations. The proposal suggests that standard INT8 integer arithmetic, combined with a rotational conversion technique, can achieve results comparable to FP8 floating-point precision. The post attracted minimal engagement on Hacker News, receiving only 3 points and 2 comments at the time of aggregation. FP8 has been increasingly adopted in AI hardware and training pipelines to reduce memory and computational costs. If validated, the INT8 ConvRot approach could offer a simpler or more hardware-compatible path for low-precision inference and training.