RTX 6000 Blackwell faces deterministic GSP kernel panics under heavy GEMM workloads
Sentdex · x · 2026-08-31
Users report that the NVIDIA RTX PRO 6000 Blackwell encounters deterministic GSP kernel panics (Xid 120 error) under sustained FP32 GEMM workloads in PyTorch/cuBLAS on Ubuntu, requiring a reboot to recover. The crash logs point to a specific address in the firmware code path across independent runs, indicating a firmware bug rather than physical hardware degradation. The issue has been reproduced across different kernel and driver versions.
More from Infra
- Ming-Chi Kuo: Nvidia Revives Rubin CPX, Production Set for Q1 2027 — zephyr_z9 · 2026-08-31
- Bun WebView API Test: Agent-Built Service Uses 100MB RAM — JeremyCMorgan · 2026-08-31
- Voibe Launches Privacy-First Speech-to-Text API for Agents — ayushtweetshere · 2026-08-31
- Macs Called the Quietest AI Infrastructure Flex Right Now — eyishazyer · 2026-08-31
- Together AI Partners With Saudi Arabia on 250MW Data Center to Avoid US Backlash — togethercompute · 2026-08-31
- Together AI: Power is the real bottleneck in AI right now — togethercompute · 2026-08-31