# RTX 4090 Tensor Unlock Doubles FP16 Throughput on Linux, and Ampere Cards Gain Up to 80% > Source: > Published: 2026-10-25 10:10:55+00:00 **A pair of open-source Linux kernel modules lift the half-rate cap Nvidia puts on FP32-accumulate tensor math in GeForce cards, taking an RTX 4090 from 165.6 to 331.6 TFLOPS in FP16. Games see nothing. Local AI workloads see a lot.** Nvidia has never hidden that GeForce cards run some tensor operations at half speed. Its own Ada Lovelace whitepaper lists the RTX 4090 at roughly 165 TFLOPS for FP16 matrix math with FP32 accumulation, half of what the same AD102 die manages when it accumulates in FP16. Workstation cards built on that silicon do not carry the penalty. The hardware is identical; the limit lives in firmware. Now an RTX 4090 tensor unlock published on GitHub removes it, and the measured gain lands almost exactly on the 2x you would expect. ## What the Ada module does The project, [gpu-unlock-ada](https://github.com/JustAResearcher/gpu-unlock-ada) by a developer going by JustAResearcher, is a Linux kernel module that loads before Nvidia’s driver. It runs a firmware transaction on the GPU that flips the SM speed-select setting, triggers a PCIe function-level reset, and then steps aside so the stock, unmodified Nvidia driver can bind to the card as normal. There is no VBIOS flash and no patched driver. The firmware is generated on the user’s own machine from pinned Nvidia inputs that are checked against SHA-256 hashes, so the repository itself ships no Nvidia binaries. On the author’s RTX 4090, FP16 with FP32 accumulation went from 165.6 to 331.6 TFLOPS, and BF16 from 165.5 to 330.6 TFLOPS. An experimental RTX 4070 Ti SUPER target, validated on a single card at a 240 W cap, doubled from 95.0 to 189.9 TFLOPS in FP16 and roughly doubled in TF32 as well. INT8 and FP8 did not move, which fits the idea that only the FP32-accumulate path was ever throttled. ## Ampere got there first The Ada work builds on [gpu-unlock](https://github.com/Zagitta/gpu-unlock) from Zagitta, which does the same job for GA102, GA103, GA104 and GA106 cards. Its RTX 3090 Ti numbers, taken from dense 8192-cube matrix multiplications at a 450 W limit, show BF16 climbing from 87.5 to 157.4 TFLOPS (+79.9%), TF32 from 42.6 to 82.6 TFLOPS, and FP16 by a smaller 35%. As [The PC Enthusiast](https://thepcenthusiast.com/nvidia-rtx-30-gpu-unlock-mod-ai-performance/) reports, community testers have since posted similar BF16 gains on RTX 3060, 3070 and 3080 cards, and a ComfyUI run on a 3090 Ti cut a BF16 image-generation job from 72 to 45 seconds. [TechPowerUp](https://www.techpowerup.com/354110/geforce-unlocked-modders-boost-rtx-20-rtx-30-and-rtx-40-compute-performance-by-up-to-2x) rounded up both projects earlier today. ## Read the fine print The Ada README is unusually candid. Its figures are “register-resident issue rates, not application GEMM or inference speedups,” and independent reproduction is still listed as pending. The 4090 result comes from one card and one cold power-on. The module has to load before anything else touches the GPU, the first attempt needs a full power-off rather than a warm reboot, multi-GPU setups and display handoff are unvalidated, and boot entries need re-checking after every kernel upgrade. Other Ada chips, A-series workstation cards and datacenter parts are refused outright. Do not expect a single extra frame, either. Games do not lean on FP32-accumulate tensor throughput, and token generation in local LLMs is bound by memory bandwidth rather than math, so gains there are small; prompt prefill reportedly picks up 14 to 36%. Where this matters is BF16 diffusion, fine-tuning and training, the exact jobs people buy 24 GB GeForce cards to do on the cheap. That is also why Nvidia is unlikely to shrug it off. The half-rate cap is one of the few hard lines separating a $1,599 GeForce from a workstation card that costs several times as much on the same die, and a firmware-level bypass is the kind of thing that tends to get closed in a later driver or firmware revision. Anyone who builds a workflow around it should keep a stock boot entry handy. For now it is Linux-only, GPL-2.0 licensed, and entirely at your own risk.