DEDUPED Hot AI repos, turned into use cases Get the Monday 5

How to unlock full-rate FP16 and BF16 tensor math on an RTX 4090 with GPU Unlock Ada

Build the kernel module, add one boot option, cold boot, and check the kernel log for the unlock.

Short answer: Run git clone https://github.com/JustAResearcher/gpu-unlock-ada.git and cd gpu-unlock-ada; then run make install to build the package locally, which downloads and verifies the pinned NVIDIA inputs and generates the firmware on your machine; Find the card's full PCI address with lspci -D and add gpu_unlock.targets= with that exact address to the kernel command line. All steps.

Reverse engineering. The unlock runs a payload derived from published analysis of NVIDIA's debug booter and changes GPU register state from a kernel module before the driver loads. It is unofficial, unsupported by NVIDIA, may void your warranty, and has been validated on very few cards.

Who it is forAn ML engineer with an RTX 4090 Linux box who wants more FP16 or BF16 throughput for training or inference.

The problemGeForce cards run FP16 and BF16 tensor math with FP32 accumulation at half rate.

The outcomeA card that reports full-rate tensor math after boot, with the stock NVIDIA driver loaded on top.

What you need first

  • A Linux machine you fully control, with a spare stock boot entry to fall back to
  • RTX 4090 (10de:2684); RTX 4070 Ti SUPER (10de:2705) and RTX 4060 Ti 16GB (10de:2805) are experimental and validated on one card each
  • Kernel headers and a working DKMS toolchain, with CONFIG_CRYPTO_LIB_SHA256 in the kernel
  • The stock NVIDIA driver
  • Secure Boot turned off, or the out-of-tree module signed with your own key
  • Your own stock benchmark numbers taken before you enable it

Steps

  1. Run git clone https://github.com/JustAResearcher/gpu-unlock-ada.git and cd gpu-unlock-ada
  2. Run make install to build the package locally, which downloads and verifies the pinned NVIDIA inputs and generates the firmware on your machine
  3. Find the card's full PCI address with lspci -D and add gpu_unlock.targets= with that exact address to the kernel command line
  4. Power the machine off and start it again for the first attempt, since a warm reboot may not give the card a fresh state
  5. Run journalctl -k -b | grep gpu_unlock and look for that card's gpu_unlock: unlocked line; a loaded module alone does not prove it worked

Prompt for your agent

On my Linux machine, build and install JustAResearcher/gpu-unlock-ada with `make install`, add gpu_unlock.targets=<my RTX 4090 PCI address> to the kernel command line, keep a stock boot entry, and tell me when to cold boot. After boot, report the gpu_unlock lines from the kernel log and whether `gpu_unlock: unlocked` appears for that card.

Caveat. The 2x figure is a register-level instruction rate; one matched run found real cuBLAS GEMM gains of 12 to 34 percent for FP16, so measure your own workload before trusting any headline number.

Use this in your AI agent

Ask your agent

Using the Deduped MCP, get the setup steps for JustAResearcher/gpu-unlock-ada and "Unlock full-rate FP16 and BF16 tensor math on an RTX 4090"

Add Deduped to Claude Code once

claude mcp add --transport http deduped https://dedupedai.com/mcp

Cursor, ChatGPT, others

Share: X LinkedIn Hacker News Email Copy link

Third-party repos we don't own. AI-written use cases can be wrong. Disclaimer