Issues

• MI350P Data

  SemiAnalysisAI/InferenceX#5010 · Filter by author volkermauel opened on Nov 11, 2026

• OperatorX: H200 dp>1 MoE rows show more than 100% of peak bandwidth; dp ranks likely time the same tokens / OperatorX：H200 上 dp>1 的 MoE 数据行显示超过 100% 峰值带宽，各 dp rank 可能在计时同一批 token

  SemiAnalysisAI/InferenceX#4996 · Filter by author positronAI opened on Nov 11, 2026
  1

• AgentX: aiperf client stalls ~300 s into profiling, then sends the backlog at once / AgentX：aiperf 客户端在测量约 300 秒时卡住，随后积压请求一次性发出

  SemiAnalysisAI/InferenceX#4995 · Filter by author xiayuqing0622 opened on Nov 11, 2026

• ubenchX/l2_distance/nvidia: make measures only GPU-count SMs and does not run on B300 / ubenchX/l2_distance/nvidia：make 只测量 GPU 数量个 SM，且无法在 B300 上运行

  SemiAnalysisAI/InferenceX#4947 · Filter by author 0z5a opened on Nov 9, 2026
  3

• AgentX warmup aborts the whole run when a request hits a keep-alive connection the server just closed

  SemiAnalysisAI/InferenceX#4533 · Filter by author zoroyihan7 opened on Oct 26, 2026

• agentic client: venv rebuilt on every run and caller's UV_CACHE_DIR overridden

  SemiAnalysisAI/InferenceX#4234 · Filter by author chaojhou opened on Oct 17, 2026

• [PowerX] centralize DCGM profiles and backfill power coverage / 集中配置 DCGM profiles 并补齐功耗覆盖

  SemiAnalysisAI/InferenceX#3592 · Filter by author edwingao28 opened on Sep 29, 2026

• [Power] Power measurement roadmap

  SemiAnalysisAI/InferenceX#2681 · Filter by author edwingao28 opened on Aug 19, 2026