Study Finds Uniform INT4 Outperforms NVFP4 on Real Gradient Tensors
A researcher comparing four-bit quantization palettes tested uniform INT4 against NVIDIA's NVFP4 format on 45 real gradient tensors from neural network training. Uniform INT4 achieved lower quantization error on 44 of 45 rotated blocks and 41 of 45 unrotated blocks, outperforming NVFP4 without any specialized format machinery. An earlier hypothesis — that INT4 only wins after a Hadamard rotation removes outliers — turned out to be an artifact of how synthetic heavy-tailed test data was generated. Optimized palette search via coordinate descent reduced error further, by up to 28%, and both biased and unbiased optimization objectives converged on nearly the same 16-value grid. The findings suggest palette design affects the magnitude of quantization error but cannot eliminate the fundamental trade-off between unbiased and low-error rounding at 16 discrete levels.
This is an AI-generated summary. ShortSingh links to the original source for the complete article.
Discussion (0)
Log in to join the discussion and vote.
Log in