gpu-unlock-ada: Enable Tensor Speed-Select on Ada GPUs

Overview

gpu-unlock-ada is a Linux kernel module that integrates the AD102 payload from TOTORONG’s AD103_win_unlocker into the Linux module and firmware-generation base from Zagitta’s gpu-unlock. It enables an SM speed-select override before the stock NVIDIA driver binds, addressing reduced tensor instruction throughput on supported GPUs without requiring a patched NVIDIA driver or UEFI application.

The primary target is the RTX 4090; other Ada GPU IDs are available only as explicit experimental opt-ins, with evidence varying by model. This is a hardware-level change, not a general-purpose GPU performance tool: measured instruction throughput gains do not necessarily translate to application gains. The README also describes a separate Windows reference port for one MSI RTX 4070 Ti, but that does not add Windows support to the Linux module.

Key Features

• Runs as a Linux kernel module early in boot, before the NVIDIA driver binds.
• Selects the AD102 firmware and payload automatically for RTX 4090 devices.
• Supports selected experimental Ada IDs through individual opt-in flags and, for some models, exact PCI-address selection.
• Builds firmware assets locally from pinned NVIDIA inputs and verifies asset hashes before use.
• Checks the payload completion marker, permissions and speed-select readout, including after PCIe function-level reset.
• Includes host-side tests and optional CUDA instruction benchmarks, plus published hardware reports and testing guidance.
• Provides package and initramfs integration for mkinitcpio, initramfs-tools and dracut.

Use Cases

• RTX 4090 owners who want to test higher FP16 or BF16 tensor instruction throughput on Linux and can validate the result on their specific card.
• Owners of listed experimental Ada GPUs who are prepared to opt in by exact device and assess the model-specific evidence before deployment.
• GPU researchers investigating tensor instruction rates and firmware-level speed selection, rather than expecting equivalent gains in every workload.
• Linux system builders who want an early-boot module and locally generated firmware assets instead of a patched NVIDIA driver or UEFI application.

Getting Started

git clone https://github.com/JustAResearcher/gpu-unlock-ada.git
cd gpu-unlock-ada
make install

Review the README’s supported-target table and boot instructions before enabling the module. Build inputs and firmware are generated locally; consult the README for distribution-specific targets, kernel command-line options and validation steps.

Considerations

The repository is very new and changing quickly, although it has recent releases, frequent commits and issue and pull-request activity. Contributions are not dominated by one person, but the short project history limits what that says about long-term maintenance or hardware coverage.

An unlock requires a compatible GPU, Linux kernel headers, a working DKMS toolchain, locally generated firmware assets and correct initramfs ordering. The first attempt requires a cold start, and a warm reboot may not restore the hardware state needed to rerun the firmware sequence. Experimental Ada support is explicitly limited to listed IDs and small numbers of tested cards; laptop results in particular may not carry across models. Reported instruction-rate gains do not ensure faster real workloads, and some measured workloads showed no benefit. The module, packaging and payload integration are GPL-2.0-only; NVIDIA firmware remains subject to its own terms.