GPUUnlock

Which cards the SM speed-select override accepts, and what it measured

Three upstream kernel modules unlock the tensor throughput NVIDIA reserves for other parts of the chip. This site is the evidence side of the project: the accepted device IDs, how well each card is actually known, and the before-and-after numbers measured on the same card. How to build and install the module is on gpu-unlock.com.

72PCI IDs accepted

Across AD102, GA102, GA103, GA104 and GA106 from the merged Ampere/Ada module, plus TU102, TU104 and TU106 from the separate Turing module. Acceptance is not the same as a confirmed card — see the card list.

3cards with published same-card numbers

RTX 4090, RTX 4070 Ti SUPER and RTX 2080 Ti. Everything else rests on user reports or is untested.

2×peak tensor gain

f16/f32 and bf16/f32 on RTX 4090. FP32 and INT8 were never rate-limited and do not move.

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 stock state. What it changes is the tensor rate limit; the core SM clock and the raster pipeline are untouched.

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 read the Turing module's own warning about sustained tensor power draw before you enable it on an RTX 20 series 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.

CardModeStockUnlockedGain
RTX 4090f16/f32165.6330.82.00×
RTX 4090bf16/f32165.5330.32.00×
RTX 4070 Ti SUPERf16/f3295.0189.01.99×
RTX 4070 Ti SUPERtf3247.594.51.99×
RTX 2080 Tif16/f3250.368.31.36×
RTX 2080 Tif16/f3255.791.41.64×
RTX 2080 Tif16/f3253.781.71.52×
RTX 2080 Tif16/f3252.877.91.48×

All benchmark rows, including the paths that did not change

Cards

Grouped by what is actually known, not by what the module accepts.

All 15 cards, including untested and unsupported IDs

Where to start

  1. Find your card in the accepted list, and see how well it is known
  2. Read the same-card numbers for the cards that were actually measured
  3. Read the install order on gpu-unlock.com — the initramfs step is not optional

Upstream projects

This site documents these repositories; it does not fork them.