Lab · SEP 17 2026
7 ms on a GPU that never went quiet
- Machine
- RTX 5090 (rented)
- Frame time, baseline → candidate
- 89.43 → 82.16 ms, −7.27 ms (−8.1%), candidate ahead 3 of 3 rounds
- Encode time, baseline → candidate
- 53.57 → 48.22 ms, −5.35 ms (−10.0%), 3 of 3 rounds
- Output records bit-identical
- 80 of 80
- im2col, per encode, in the per-kernel GPU trace
- 2.32 → 1.87 ms
Vectorising the float16 conversions in a CUDA port of SAM 3 took 7.27 ms (8.1%) off a frame, bit-identical, on a rented GPU that failed its idle check on every single run.