Unlock NVIDIA SM speed-select, with the numbers attached
Three upstream kernel modules let consumer GeForce cards reach tensor throughput NVIDIA reserves for other parts. This site collects what they support, what was measured on the same card before and after, and what breaks.
What the unlock does
The module enables the SM speed-select override before the stock NVIDIA driver binds. It runs from the initramfs, so the driver never sees the card in its locked state. That ordering is the whole mechanism — loading the module after the driver has bound cannot work.
Hardware risk is real and documented
The unlock writes speed-select registers and performs an FLR before the driver loads. Keep an unmodified stock boot entry, validate one cold boot before making anything persistent, and treat the AD103 path as experimental — it was validated on a single card.
Same-card measurements
Every row below compares stock and unlocked on the same card with the same binary. Modes that were already at full rate are omitted here and shown in full on the benchmarks page.
| Card | Mode | Stock | Unlocked | Gain |
|---|---|---|---|---|
| RTX 4090 | f16/f32 | 165.6 | 330.8 | 2.00× |
| RTX 4090 | bf16/f32 | 165.5 | 330.3 | 2.00× |
| RTX 4070 Ti SUPER | f16/f32 | 95.0 | 189.0 | 1.99× |
| RTX 4070 Ti SUPER | tf32 | 47.5 | 94.5 | 1.99× |
| RTX 2080 Ti | f16/f32 1024³ | 50.3 | 68.3 | 1.36× |
| RTX 2080 Ti | f16/f32 2048³ | 55.7 | 91.4 | 1.64× |
| RTX 2080 Ti | f16/f32 4096³ | 53.7 | 81.7 | 1.52× |
| RTX 2080 Ti | f16/f32 8192³ | 52.8 | 77.9 | 1.48× |
Cards
Grouped by what is actually known, not by what the module accepts.
- GeForce RTX 409010de:2684AD102
- GeForce RTX 4070 Ti SUPER10de:2705AD103
- GeForce RTX 2080 Ti10de:1e07TU102
- GeForce RTX 3090 Ti10de:2203GA102
- GeForce RTX 309010de:2204GA102
- GeForce RTX 3080 10 GB10de:2206GA102
- GeForce RTX 3080 20 GB (memory upgrade)10de:2206GA102
- GeForce RTX 307010de:2484GA104
Where to start
- Check that your card is in the accepted list, and how well it is known
- Read the install order — the initramfs step is not optional
- Read the failure modes before you flash anything
Upstream projects
This site documents these repositories; it does not fork them.
- Zagitta/gpu-unlock — the Ampere module: GA102, GA103, GA104, GA106 (v1.0.0)
- JustAResearcher/gpu-unlock-ada — the Ada fork: adds AD102 RTX 4090 and an experimental AD103 canary (v1.1.1)
- nlzy/turing-unlock — the Turing module: TU102, TU104, TU106 (RTX 20 series); only the 2080 Ti is measured
- miounet11/gpu-unlock — the public usage notes for this site