A terminal monitor that names the local models sitting in your video memory
Emerged finds new products worth a look, daily: what’s new today
gpu-unlock
A Linux kernel module that doubles half-precision matrix speed on RTX 30 cards
Is this yours? Claim itAnyone running local inference or fine-tuning on a 3090 can reach close to professional-card tensor throughput for the price of a reboot, with the NVIDIA driver left untouched.
The module writes the card's speed-select registers and resets the GPU before NVIDIA's driver claims it, so nothing in the driver stack is patched and updates do not undo it. Install is a clone and make install, with Arch, Debian and Fedora packaging. Users report dense BF16 GEMM gains of 25 to 85 percent depending on card and power limit, output bit-identical, while INT8 and decode work is unchanged. GPL-2.0, 8.7k stars, 29 commits, 131 open issues. The display blanks until the driver takes over, hibernation is unhandled, and the firmware stays under NVIDIA's terms.
- Open source
More like this
Share one NVIDIA GPU between the host and several KVM guests, no passthrough
GPU specs and live rental prices from 24 clouds, side by side
Write PyTorch on your laptop while the CUDA work runs on a remote NVIDIA box
A Linux desktop built on a Vulkan renderer of its own
A slimmed MiniMax-H3 that makes video with sound on a single 24 GB card