● Covered by 2 sources · 2 reportsMedium impact2 neutral
Linux Kernel Module Unlocks Nvidia 3000 Series FP/BF16 Performance
Key points
- Kernel module unlocks FP/BF16 instruction rate on Nvidia 3000 series cards.
- Doubles instruction issuing speed for FP/BF16 FMA into FP32 operations.
- Improves performance for LLM prefill, diffusion, and FlashAttention.
- Does not require patching NVIDIA drivers with each new version.
- UEFI tool nv-uefi-unlocker unlocks RTX 20, 30, and 40-series GPUs.
- The tool runs from a FAT32 USB stick before the Nvidia driver loads.
- A laptop RTX 4060 showed 65% higher FP16 math performance.
- The tool uses community V67 payload with read-back verification.
- The unlocker is now available for Windows users.
- The release ships scripts and self-built binaries, not Nvidia firmware.
Kernel Module Functionality
A new Linux kernel module has been developed to unlock the FP/BF16 instruction issue rate on Nvidia 3000 series graphics cards. This module combines the CMP unlocker's V67/PLM technique with pre-driver timing from the UEFI approach. It executes the unlock, writes speed selectors (SS0 0x88888888 and SS1 0x8IMLA4), and performs FLR before the unmodified NVIDIA driver binds to the card. This method avoids the need for a UEFI application or a patched NVIDIA driver.
Performance Impact
The unlock doubles the instruction issuing speed of FP/BF16 FMA into FP32 operations on GA102, GA103, GA104, and GA106 cards. User reports indicate significant gains in dense GEMM throughput. For example, LLM prefill shows improvements of 14–36%, and diffusion and FlashAttention also benefit. LLM decode and INT8 work, which are memory bandwidth limited, do not see similar gains. Outputs remain bit-identical before and after the unlock.
Installation and Compatibility
The module is installed by cloning the repository, running 'make install', and then appending 'gpu_unlock.targets=all' to the kernel command line. This can be done via GRUB for Debian/Ubuntu/Mint, Arch/CachyOS, Fedora/RHEL, systemd-boot, or UKI/generated entries. The module supports various GeForce IDs, and users are encouraged to report compatibility for their specific cards. The build process downloads pinned NVIDIA/CachyOS inputs, checks SHA-256, and generates firmware in the build directory and binary package.
Updates
- UEFI tool nv-uefi-unlocker unlocks RTX 20, 30, and 40-series GPUs.
- The tool runs from a FAT32 USB stick before the Nvidia driver loads.
- A laptop RTX 4060 showed 65% higher FP16 math performance.
- The tool uses community V67 payload with read-back verification.
- The unlocker is now available for Windows users.
- The release ships scripts and self-built binaries, not Nvidia firmware.
✨ This summary was generated by AI from the outlets' reporting listed below. It is not independently verified and may contain errors — check the original sources. How BrevFeed works →
The daily brief
One email each morning: the day's tech stories, clustered across outlets and summarized. No account needed.
One email a day. Unsubscribe in one click, any time.
Today's brief
Spend a few minutes, get the whole day. Every topic's top stories in one hands-free rundown — listen, watch, or read the transcript.
▶ Play today's brief
New every morning, and the back catalogue is archived by date.
Primary sources
GitHub cmc_internal/api GitHub github/collect GitHub Zagitta/gpu-unlock GitHub _private/browser GitHub get-started/accessibility GitHub open-source/sponsorsHow outlets covered it
A new open-source UEFI tool, nv-uefi-unlocker, boots from USB to remove Nvidia's half-speed AI throttle on RTX 20, 30, and 40-series GPUs. This allows tensor units to run at full rate, offering increased FP16 performance without VBIOS flashes or patched drivers, and is now available for Windows users.
A new Linux kernel module allows Nvidia 3000 series GPUs to achieve the FP/BF16 instruction issue rates of their professional equivalents. This module doubles the instruction issuing speed of FP/BF16 FMA into FP32 operations on GA102, GA103, GA104, and GA106 cards, significantly improving performance for matrix-heavy workloads like LLM prefill and diffusion.