Three-line fix removes noisy math on cheap Tesla P100 GPUs
automatically uses FP16 calculations on s that can run them quickly. The GTX 10 series and P40 were excluded from this behavior long ago, but the Tesla P100 was not because it has fast FP16 . That choice can add needless numerical error to calculations on the P100 despite improving speed.
The correction changes only three lines of code and is already included in 0.3.0 and two related forks. A request to fix the original project was also filed with GGML as issue 25593.
Key points
- The Tesla P100 can perform fast FP16 calculations, but 's existing choice may introduce needless numerical error.
- The correction requires a three-line code change.
- The fix is available in 0.3.0.
- The change has been merged into two forks.
- The original GGML project has a fix request filed as issue 25593.