GPU-unlock: FP/BF16 instruction issue rate on Nvidia 3000's cards

Unlock FP/BF16 instruction issue rate on Nvidia 3000's cards to their professional equivalents

This Linux kernel module combines the CMP unlocker's V67/PLM technique with the pre-driver timing of the UEFI approach. It runs the unlock, writes the speed selectors (SS0 0x88888888, which selects the full BF16 rate, and the 4-bit SS1 0x8 IMLA4) and performs FLR before the unmodified NVIDIA driver binds the card. It is neither a UEFI application nor a patched NVIDIA driver.

The benefit of this early kmod driver is that it does not require manual patching each time nvidia releases a new driver version.

The effect of this unlock is that it doubles the instruction issuing speed of FP/BF16 FMA into FP32 operations on GA102, GA103, GA104 and GA106 cards.

3090 Ti dense 8192³ GEMM at 450 W, five-sample medians with FP32 accumulation:

Results reported by users. The module accepts every GeForce ID listed in kernel/gpu_unlock/main.c, but the cards below are the ones someone has reported on. Please add yours with a compatibility report, whether it worked or not.

The gains are dense GEMM throughput from each report's own benchmark, so they are not directly comparable. The power limit matters a lot: the lowest figure (+25%) is from cards capped at 225 W.

Matrix-heavy work speeds up: LLM prefill (+14–36% in #9, #15, #25 and #33), diffusion and FlashAttention. LLM decode, which is limited by memory bandwidth, and INT8 work stay about the same. The reports that compared outputs found them bit-identical before and after.

Reports also cover up to four cards in one machine, the card driving the display, and cards passed through to a VM with vfio-pci (#17).

First run:

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

Then append gpu_unlock.targets=all to your kernel command line, keeping existing options:

• GRUB (Debian/Ubuntu/Mint): add it inside GRUB_CMDLINE_LINUX_DEFAULT in /etc/default/grub, then run sudo update-grub.
• GRUB (Arch/CachyOS): same edit, then run sudo grub-mkconfig -o /boot/grub/grub.cfg.
• Fedora/RHEL: run sudo grubby --update-kernel=ALL --args="gpu_unlock.targets=all".
• systemd-boot: append it to the options line in your boot entry (bootctl status shows its location).
• UKI / generated entries: add it to /etc/kernel/cmdline and regenerate your boot images/entries (e.g. sudo mkinitcpio -P for mkinitcpio UKIs).

Every build downloads the pinned NVIDIA/CachyOS inputs over HTTPS, checks their SHA-256, and generates firmware only in the build directory and the binary package. All three formats install through the same make install-* targets. Packages are built from the committed HEAD.

Build, install build dependencies, and install the packages in one step (asks for your sudo password):

make install        # detect the distribution and select an installer
make install-arch   # makepkg -sif
make install-deb    # apt-get build-dep, dpkg-buildpackage, apt-get install
make install-rpm    # dnf builddep, rpmbuild, dnf install

make install uses /etc/os-release (ID, then ID_LIKE) to detect Arch, Debian/Ubuntu, or Fedora/RHEL/CentOS families. The RPM installer requires DNF. For other distributions, select a compatible explicit target above.

make arch, make deb and make rpm only build. Debian and RPM packages go to build/deb and build/rpm. The NVIDIA inputs are cached in ~/.cache/gpu-unlock/inputs (make inputs; override with INPUTS=). dpkg-buildpackage itself never uses the network, so a clean-chroot build (sbuild or pbuilder) needs GPU_UNLOCK_INPUTS pointing at that cache. The make dist archive, the Debian source package and the .nosrc.rpm contain no NVIDIA archives or generated firmware.

The packages put gpu_unlock and its firmware in the initramfs and load it there, before nvidia can bind a card:

• mkinitcpio: /etc/mkinitcpio.conf.d/zz-gpu-unlock.conf inserts gpu_unlock right before nvidia in MODULES, so a display card is blank for as short a time as possible. Without nvidia in the image, it goes last and still runs before an nvidia loaded from the root filesystem. Images are rebuilt on install. mkinitcpio ignores drop-ins for an image built with -c <config> or a preset _config; put gpu_unlock right before nvidia in that config's MODULES yourself.
• initramfs-tools: /usr/share/initramfs-tools/modules.d/gpu-unlock.conf.
• dracut: /usr/lib/dracut/dracut.conf.d/gpu-unlock.conf.

With initramfs-tools and dracut, the module is built for every installed kernel that has headers, and those kernels' images are rebuilt on install and again on removal. Kernels installed later get the module and a new image from the kernel's own install hooks.

There is no softdep nvidia pre: rule: kmod honours only the first softdep nvidia line it reads, so other packages' rules (nvidia-utils, gddr6_hwmon) would silently override it, or it would override theirs.

The module does nothing unless the kernel command line opts in:

gpu_unlock.targets=all                          # every unbound GA102/GA103/GA104/GA106 card
gpu_unlock.targets=0000:21:00.0                 # only these cards
gpu_unlock.targets=0000:21:00.0,0000:22:00.0    # comma-separated

Addresses are full PCI names as lspci -D prints them (hex digits in either case), independent of GPU numbering and display role. An entry that is not a GA102/GA103/GA104/GA106 card, including all mixed with addresses, is logged and ignored. The cards are unlocked concurrently, about a second in all however many there are, and each on its own: a card that is already bound or powered down is skipped. Bus mastering that the board firmware left on (Supermicro X11 does this whatever its option ROM and bus master settings) is cleared first; a card some driver already initialised is still refused. Then nvidia loads and binds every card as usual.

The display card works too. Its firmware framebuffer (the boot console and Plymouth) keeps running while the unlock's firmware steps run, and is removed just before the card's reset, as nvidia-drm would do a moment later anyway. Its screen is blank from that reset until nvidia-drm takes over.

Loading the module by hand after boot does nothing to a card that nvidia has bound.

Suspend to RAM powers the cards off, which resets the speed selectors. After resume the module unlocks every card it unlocked at boot again, before nvidia resumes it: its suspend notifier runs just ahead of the one nvidia resumes its cards from (NVreg_UseKernelSuspendNotifiers=1, the default in current drivers; nvidia-resume.service runs later still). This adds about a second to resume. A card that kept power, as in suspend to idle, is still unlocked and is only read. A card that nvidia has already resumed, or that is bound to another driver, is left as it is, and the log says why. Add gpu_unlock.resume=0 to the kernel command line, or write 0 to /sys/module/gpu_unlock/parameters/resume, to turn this off. Hibernation is not covered.

Check the result with:

journalctl -k -b | grep gpu_unlock
lspci -nnk -s 0000:21:00.0

To stop unlocking, remove gpu_unlock.targets= from the kernel command line; for a one-off boot, delete it in the bootloader's edit screen. Removing the packages also removes the module and the initramfs drop-ins.

If the firmware on a card never stops, the module logs it and keeps the DMA buffers that firmware may still use until the next power off. Power off before trying again.

The kernel module and local packaging are GPL-2.0-only. The generated payload derives from the pinned CMP reference; NVIDIA firmware remains under its own terms. See firmware/NOTICE.md and licenses/ for attribution and notices.

Help me buy more GPUs <3

https://buymeacoffee.com/zagitta

ETH: 0xf4eee937937f92038D70B16AbD7736D8a53a1c5c

BTC: bc1qwpkwp70fn76z3s4kg2l9y3j4kst95twcznuugf

SOL: 8Rih6HY7V4LcdjYHz4iSffT4fTniqw5CACaBigSVxNYu