Unverified50% confidenceBenchmarkExact time
The default NCCL reduce-scatter can slow down a GEMM that originally completes in a single wave by 69%.
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
10/30/2026
First Seen
Valid until: 1/28/2027
Sources
Related Entities
Related Claims
UnverifiedFP8 has a narrower numerical range and lower precision, and correct usage requires fine-grained scaling factor management and numerical stability guarantees at the GEMM kernel level.67% similarUnverifiedTIDE's reduce-scatter SM time is reduced to 1/12.5 to 1/14.5 of the original, and its interference with concurrent computation is reduced by up to an order of magnitude.62% similarUnverifiedCompared to pure LLM solutions, RAG significantly reduces the occurrence of hallucination issues.56% similarUnverifiedTIDE reduces step time by up to 6.0% compared with FSDP2+SymmMem.56% similarUnverifiedUnder the condition of reaching the same stationarity tolerance, the schedule optimization time of PB-AMD is lower than that of projected gradient descent.56% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1114841API
curl https://kongchang.com/api/v1/knowledge/claims/1114841MCP
get_claim(id=1114841)