Understanding FP8 Numerical Spaces in LLM Quantization
FP8 is becoming increasingly important in modern LLM systems, but its behavior is often difficult to reason about intuitively. Unlike integer quantization, FP8 combines floating-point’s non-uniform numerical spacing with quantization scaling challenges. In this talk, we will explore FP8 formats, outlier problems, and scaling effects through visualization, with the goal of building stronger intuition for low-precision floating-point quantization in LLMs.