Unverified50% confidenceFactExact time
In collective communication (such as All-Reduce), all GPUs need to exchange gradients synchronously, and microsecond-level jitter on any single link puts the entire group of GPUs into a wait, creating a bucket effect.
1
Sources
50%
Confidence
Long-term
Relevance
10/26/2026
First Seen
Sources
Related Entities
Related Claims
Unverified张量并行通过 All-Reduce 操作在所有参与 GPU 之间同步中间激活值，通信量随并行度线性增长79% similarUnverified当同一Warp内不同线程执行不同分支路径时，GPU必须串行执行所有分支，造成线程束分支发散导致的性能损失72% similarUnverified智能体工作流的GPU负载呈现突发性和不规律性，形成突发-静默交替的复杂负载曲线72% similarUnverifiedThrough overlapping data transfer and computation (double buffering/pipeline prefetching), the GPU computes the current batch of experts while simultaneously prefetching the next batch of expert weights in the background, converting serial waiting into pipelined parallelism.67% similarUnverifiedDynamic batching allows the system to dynamically combine requests into batches within an extremely short time window, finding the optimal balance between latency and GPU utilization, which is superior to static batching that requires waiting to accumulate a fixed number of requests.66% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1099954API
curl https://kongchang.com/api/v1/knowledge/claims/1099954MCP
get_claim(id=1099954)