Clustered CUDA launch runs 128 blocks when gridDim says 64: a hidden driver bug
mike64_t · x · 2026-10-23
While testing a CUDA kernel, the author noticed something odd: a plain launch ran 64 blocks as expected, but with thread block clustered launch and device update enabled, 128 blocks actually ran while gridDim.x still reported 64 — with no errors raised.
The follow-up pinpoints the cause: c[0][0x0c] and CTARASTERWIDTH were not kept in sync, i.e. the driver-filled kernel parameter mismatched the raster width constant under clustered launch, so more CTAs executed than intended. A deep GPU driver/launch debugging find — takeaway: under clustered launch, don't trust gridDim; verify actual block execution counts.
More from Fun
- The classic tech joke: 'My servers were NOT impacted by the RCE from the PHP tech debt' — MickeySteamboat · 2026-10-23
- CTO exits blood-testing startup of Elizabeth Holmes' partner, dubbed "Theranos 2.0" — annatonger · 2026-10-23
- AI assistant Boba dissects 'friend' marketing: 'the warmth is the growth strategy' — gharik · 2026-10-23
- VC Parody mocks the YC founder archetype: 'taste is the only moat in AI slop' — DesireeCachette · 2026-10-23
- Can't automate a job that's 90% meetings about meetings — StewartalsopIII · 2026-10-23
- Polymarket AI prediction crashes to 47% after weeks near 90%, $849K volume — patience_cave · 2026-10-23