A Linux kernel module that applies an SM speed-select override before the stock NVIDIA driver binds, roughly doubling FP16/BF16 tensor throughput with FP32 accumulation on RTX 4090 (AD102) and experimentally on RTX 4070 Ti SUPER (AD103).
Useful for
- Build and install the module locally via make install for Arch, Debian or RPM.
- Enable the unlock at boot by passing the card's PCI address in gpu_unlock.targets.
- Measure FP16/BF16 tensor throughput gains on your RTX 4090 after a cold boot.
Before using
- License: GPL-2.0
- Installation and requirements have not been verified; review the README before running
Why now
- Created on GitHub
- 2026-10-23 UTC
- First seen by GitNova
- 2026-10-24 UTC
- Magnitude
- 8.6 / 10 · growing interest
Only 1 complete days of history; a reliable usual pace is not established. On 2026-10-24: +1,241 stars (UTC)
Maximum over those 1 complete days: +1,241 on 2026-10-24; the current measurement has not exceeded it (UTC)
Measured: October 25, 2026, 06:26 UTC
Measurement details
- Created on GitHub: 2026-10-23; first seen by GitNova: 2026-10-24 (UTC)
- +490 stars measured 2026-10-25, 00:09–06:17 UTC
- Only 1 complete days of history; a reliable usual pace is not established. On 2026-10-24: +1,241 stars (UTC)
- Maximum over those 1 complete days: +1,241 on 2026-10-24; the current measurement has not exceeded it (UTC)
- Magnitude 8.6/10