Unverified50% confidenceFactExact time
NVFP4 has very few mantissa bits, and the rounding error of a single matrix multiplication is significantly larger than that of BF16, but block scaling keeps its gradient estimates unbiased in a statistical sense.
1
Sources
50%
Confidence
Long-term
Relevance
11/10/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedThe degree-3 and degree-4 polynomial coefficients in the paper are obtained by minimizing the maximum error (minimax approximation) in the BF16/FP16 target format.75% similarUnverifiedNVFP4量化方案相比BF16基线实现了1.5倍的推理加速,同时在top-1准确率上保留了92%到97%的水平72% similarVerifiedINT4 quantization compresses weights that were originally represented by 65,536 distinct values (FP16) down to only 16 discrete levels.72% similarUnverifiedUnder BF16, if the result of any floating-point operation cannot be exactly represented in BF16, it must be rounded to the nearest representable value, producing a relative error on the order of about 2^-8.72% similarUnverifiedRRF uses only ranks rather than raw scores, avoiding the problem that the dimensions of dense vectors and sparse BM25 are not comparable.71% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1154474API
curl https://kongchang.com/api/v1/knowledge/claims/1154474MCP
get_claim(id=1154474)