An open-source kernel module just gave NVIDIA's six-year-old RTX 3090 roughly two thirds more Tensor Core throughput, and it never touches the VBIOS.
Zagitta's project gpu-unlock targets an instruction issue limit on Ampere GeForce cards for FP16 and BF16 Tensor Core math with FP32 accumulation.
Those operations run at about half the rate of the professional cards, so the module loads before the graphics driver, rewrites the speed selection registers, resets the GPU, and returns control to the stock driver.
One community report on an RTX 3090: 129.6 TFLOPS in BF16 matrix multiplication, up from 77.5.
TF32 went from 40.2 to 69.8.
On a 3090 Ti, BF16 climbed from 87.5 to 157.4 TFLOPS.
The developer reports up to 80% on certain compute benchmarks.
Picture an engineer pricing a new card this month because a fine-tune took all night.
The gains here are workload shaped.
One user saw about 16% better long context prompt processing during local inference with token generation basically flat.
ComfyUI testing on a 3090 Ti showed 47 to 59% more throughput on BF16 image workloads.
No independently verified gaming numbers exist.
The ceiling on older cards was never the silicon alone.
It was a register setting nobody had tested.
When a single developer can reopen headroom on a 2020 card, the cost of running the next experiment drops while the hardware stays exactly the same.
RTX 3090 capacity at $0.15/hr, with Buy Now Pay Later so you start at zero upfront and pay after one month.
What is the oldest GPU you still run in production, and would you put a kernel module in front of it?
https://ai.bhkcloud.com/?utm_source=linkedin&utm_medium=social&utm_campaign=bhk_social_2026w43&utm_content=morning
Originally posted on linkedin_personal