BitBOINC

For NVidia GPUs (4060Ti, 4070Super, and 4070Ti), all tasks are failing. Is there a specific Cuda version that needs to be used for these WUs? (Cuda 12 vs. Cuda 13?)

---

Thanks for the detailed report and the crash dumps — this turned out to be a real bug, not a driver fluke.

Root cause: Keraunos's GPU kangaroo kernel was doing 64 search rounds per kernel launch, and each round includes two expensive modular-inverse operations (~256-step Fermat inverse each). On Ada Lovelace cards (RTX 40-series) this pushed individual kernel launches past Windows' WDDM TDR timeout (~2 seconds by default), so the driver killed the kernel mid-flight — which is exactly the "Breakpoint Encountered (0x80000003)" exception you and two other users hit at the same instruction offset.

Fix: reduced rounds-per-launch from 64 to 8, so each launch finishes well under the TDR threshold. Deployed today for both the Windows and Linux Turing+/Ampere/Ada CUDA builds (app version 13). Could you update and let us know if it's stable now?

---

Guess I'll need to empty the jobs that I had waiting to get to the newer version (.13 to .18)

---

First 2 WUs of new .18 version successfully completed and validated! :) Took 2 hrs 10 min.

---

Nice :)

---