Unverified50% confidenceFactExact time
In large-scale GPU matrix operations, the associative law of floating-point addition does not strictly hold, and different parallel reduction orders, batch sizes, and memory scheduling can cause minor discrepancies in logits.
1
Sources
50%
Confidence
Long-term
Relevance
9/11/2026
First Seen
Sources
Related Entities
Related Claims
VerifiedIn GPU parallel reduction, because floating-point operations do not satisfy the associative law, differences in the execution order of different warps cause deviations in numerical results, which is the root cause of non-determinism.80% similarUnverifiedFlattening a matrix and feeding it to an MLP results in a non-fixed input dimension due to variable row counts, and breaks permutation invariance.72% similarUnverified浮点运算不满足结合律，不同线程束的执行顺序差异会导致数值结果出现细微偏差72% similarUnverified自回归推理时batch size通常为1，每个token生成只做一次矩阵-向量乘法，算术强度极低，导致memory-bound问题，单用户场景下GPU利用率往往不到10%69% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/901338API
curl https://kongchang.com/api/v1/knowledge/claims/901338MCP
get_claim(id=901338)