README

GPU unlock 1.1.1 — RTX 4090 and experimental Ada GPUs

New to this project? Start with the copy-paste Quick Start.

This Linux kernel module enables the SM speed-select override before the stock
NVIDIA driver starts. RTX 4090 (10de:2684, AD102) support has demonstrated
approximately 2× FP16 and BF16 tensor throughput with FP32 accumulation:

| Mode | Before | After |
|---|---:|---:|
| FP16 → FP32 accumulate | 165.6 TF/s | 331.6 TF/s |
| BF16 → FP32 accumulate | 165.5 TF/s | 330.6 TF/s |

The override survived PCIe Function Level Reset (FLR), stock NVIDIA 615.78.08
initialization and CUDA benchmarks. INT8, dp4a and scalar FP32 stayed near their
baseline rates. The original transaction preceded the repository cleanup; see
the technical report and
original evidence.

An independent owner run of revision 8855e7b also completed the RTX 4090
transaction and NVIDIA driver handoff on Linux 7.2.11-arch1-1. Matched
stock/unlocked measurements with NVIDIA 615.78.08 and the same 450 W cap
found FP16/FP32 instruction throughput 160.4 → 320.4 TF/s and BF16/FP32
159.7 → 320.2 TF/s. All exercised numerical checks passed. Real cuBLAS
GEMM median throughput varied by shape: FP16 gained 12–34%, BF16 gained
16–20% at two shapes and was essentially unchanged at 4096³. These are one
matched run pair with the desktop active, not whole-application speedups or
an extended stability test. See the matched evidence and limits
and testing workflow.

A modified 48 GB RTX 4090 (10de:2684, AD102) on Ubuntu 26.04 with Linux
7.0.0-38-generic and NVIDIA 610.57.04 open kernel modules also completed the
transaction and driver handoff on 2026-10-25: fmla32 went from 1 to 0,
FP16/FP32 instruction throughput 164.9 → 328.9 TF/s, BF16/FP32
164.8 → 328.7 TF/s, with INT8, dp4a and scalar unchanged and no Xid. A
quantized mixture-of-experts serving workload on the same card showed no
gain (8K prefill 0.983×), which is consistent with quantized matmuls not
consuming the FP32-accumulate tensor path. That is one card, one cold boot and
one workload. See the modified-card evidence and limits.
Further matched owner reproductions used NVIDIA 610.57.04
and 595.91.07.

A separate cold test of the Ada port on one RTX 4070 Ti SUPER
(10de:2705) with NVIDIA 580.178.04 measured FP16/FP32 tensor issue rate
at 95.0 → 189.9 TF/s and BF16/FP32 at 95.0 → 189.7 TF/s under the
same 240 W power cap. INT8 stayed at 379.5 TOP/s. The stock driver loaded,
the NVIDIA RM getter reported full FMLA32 rate, deterministic FP16/BF16 GEMM
outputs matched exactly, and owner mining recovered with fresh accepted shares.
A second matched instruction benchmark found TF32/FP32 47.49 → 94.51 TF/s
(1.99×), with its FP16/FP32 control also near 2×. Unsigned INT4/INT32 stayed
at approximately 754.8 TOP/s in both states. These are register-resident
issue rates, not application GEMM or inference speedups.
A matched, verified E4M3 FP8/FP32 CUTLASS GEMM measured 59.326 TF/s stock →
59.495 TF/s unlocked on this card, a 0.3% difference. Th

Top issues

Latest release

GPU Unlock 1.1.1 publishes a reproducible SM89 tensor issue-rate probe and matched stock/unlocked TF32 and unsigned INT4 results on one RTX 4070 Ti SUPER (10de:2705, NVIDIA 580.178.04, 240 W cap). The Ada unlock transaction is unchanged from 1.1.0; the module and DKMS version metadata advance to 1.1.1.

| Register-resident mode | Stock | Unlocked |
| --- | ---: | ---: |
| FP16 inputs, FP32 accumulate (control) | 94.96 TF/s | 188.97 TF/s |
| TF32 inputs, FP32 accumulate | 47.49 TF/s | 94.51 TF/s |
| Unsigned INT4 inputs, INT32 accumulate | 754.81 TOP/s | 754.68 TOP/s |

The same cubin,