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. That kernel reaches only about 17% of this card's FP8 peak, so it cannot resolve an issue-rate change. Other measurements do put FP8 in the affected group: an 8192³ E4M3 GEMM on a Windows RTX 4080 SUPER measured 199.4 → 310.4 TF/s (+55.7%) and 199.4 → 298.7 TF/s (+49.8%) in two runs (see goldlionren/gpu-unlock-4080s; its author flags FP8 variance and keeps FP8 out of that repo's own pass/fail gate), and register-resident E4M3 probes read 603.1 TF/s unlocked on the modified 48 GB RTX 4090 (see that evidence directory) and 165.45 → 323.95 TF/s (1.96×) on a Windows RTX 4070 Ti SUPER (see A-XIANGQAQ/rtx4070tis-ad103-tensor-unlock). The 4090 figure is 91% of its 660.6 TF/s full FP8 rate; NVIDIA rates a stock RTX 4090 at 330.3 TF/s for FP8 with FP32 accumulate and 660.6 with FP16 accumulate. So FP8 does move at the issue-rate level; a matched AD103 pair is 1.96×. At a fixed power limit the application GEMM stays below that: an unlocked RTX 4090 peaked at 346 W against 313 W stock and spent both MMA and GEMM passes under SW Power Cap (see the 610.57.04 evidence). This is one card; the unlock succeeded on two cold boots and two ordinary warm reboots, the last from a persistent EFI default. Throughput was measured on the first cold boot. See the 2705 canary report and its sanitized raw benchmark output.

The AD102 payload and its debug-booter analysis come from TOTORONG/AD103_win_unlocker. This fork integrates that published payload into Zagitta/gpu-unlock's Linux module, firmware generation and boot packaging. The payload is essential to the result.

Experimental Windows contribution — Latence.app

A separate MSI RTX 4070 Ti AD104 Windows UEFI reference port records one successful manual boot with NVIDIA 616.56 and matched cuBLAS FP16/FP32 +80.1%, BF16/FP32 +76.4% measurements. It is tied to one captured MSI Z490 profile; it does not add AD104 admission to the Linux module. See the report and limitations. No firmware/Windows/ROM binaries or automatic boot installer are distributed. Contribution by Latence.app; original upstream credits and licences are retained.

This is the beginning of the README. Read the rest at the original. Read the full README ↗