GPU-unlock: Enhancing FP/BF16 Instruction Rates for Nvidia 3000 Series

This project, hosted on GitHub by Zagitta, provides a method to unlock the FP/BF16 instruction issue rates on Nvidia 3000-series consumer cards, bringing them in line with their professional counterparts.

🛠️ Technical Overview

The gpu-unlock tool is a Linux kernel module (kmod) that implements a hybrid approach to unlocking. It merges the V67/PLM technique (from the CMP unlocker) with the pre-driver timing typically associated with UEFI-based methods.

How it Works

The module executes the unlock sequence before the standard NVIDIA driver binds to the hardware. The process follows this logic:

1. Unlock Execution: Initiates the unlock sequence.
2. Speed Selector Configuration:
   • Sets SS0 to 0x88888888 (enabling full BF16 rates).
   • Sets the 4-bit SS1 to 0x8 (IMLA4).
3. Reset: Performs a Function Level Reset (FLR).
4. Driver Binding: The unmodified NVIDIA driver then takes control of the card.

Key Advantage: Because this is an early-loading kernel module and not a patched NVIDIA driver, users do not need to manually re-patch their drivers every time a new NVIDIA version is released.

Boot Sequence Logic

---

🚀 Performance Impact

The primary effect is a doubling of the instruction issuing speed for FP/BF16 FMA $\rightarrow$ FP32 operations. This applies to cards utilizing the GA102, GA103, GA104, and GA106 chips.

Benchmark: RTX 3090 Ti

Dense $8192^3$ GEMM at 450 W (Median of 5 samples, FP32 accumulation)

Compatibility & User Reports

The module supports all GeForce IDs found in kernel/gpu_unlock/main.c. Below is a summary of reported results:

Note: Gains vary based on benchmarks used. Power limits are critical; for instance, the lower $+25\%$ gain was observed on cards capped at $225\text{W}$.

Real-World Application

• Accelerated: Matrix-heavy tasks such as FlashAttention, Diffusion models, and LLM prefill (increases of $+14\text{--}36\%$).
• Neutral: LLM decode (bottlenecked by memory bandwidth) and INT8 operations remain unchanged.
• Accuracy: Outputs are confirmed to be bit-identical regardless of the unlock.
• Versatility: Confirmed working with up to 4 GPUs, primary display cards, and vfio-pci VM passthrough.

---

📦 Installation Guide

Initial Setup

• Clone the repository: git clone https://github.com/Zagitta/gpu-unlock
• Enter directory: cd gpu-unlock
• Install: make install

Kernel Configuration

You must append gpu_unlock.targets=all to your kernel command line. Depending on your bootloader:

• GRUB (Debian/Ubuntu/Mint): Edit /etc/default/grub $\rightarrow$ GRUB_CMDLINE_LINUX_DEFAULT $\rightarrow$ sudo update-grub.
• GRUB (Arch/CachyOS): Edit config $\rightarrow$ sudo grub-mkconfig -o /boot/grub/grub.cfg.
• Fedora/RHEL: sudo grubby --update-kernel=ALL --args="gpu_unlock.targets=all".
• systemd-boot: Add to the options line (check bootctl status for path).
• UKI: Add to /etc/kernel/cmdline and regenerate images.

---

🛠️ Build System & Distribution

The build process is automated and secure, downloading pinned NVIDIA/CachyOS inputs via HTTPS and verifying them with SHA-256 hashes.

Installation Targets

The make install command automatically detects your OS via /etc/os-release. You can also specify targets manually:

Developer Notes:

• Inputs are cached at ~/.cache/gpu-unlock/inputs.
• To use a clean-chroot build (e.g., sbuild), set the GPU_UNLOCK_INPUTS environment variable.
• The make dist archive and source packages do not contain NVIDIA binaries or firmware.

Boot Flow & Initramfs

To minimize display downtime, the module is loaded into the initramfs before the NVIDIA driver:

• mkinitcpio: A config file at /etc/mkinitcpio.conf.d/zz-gpu-unlock.conf ensures gpu_unlock is placed immediately before nvidia in the MODULES array.
• If nvidia is not in the image, gpu_unlock loads last but still precedes any driver loaded from the root filesystem.