An RTX 2070 is rated for 36 TFLOPS of FP32-accumulation matrix math.
For years that rating looked like the hardware talking.
Partly it was software.
This week a GitHub developer going by physicsdolphin published NV UEFI Unlocker, an open source tool that changes GPU behavior inside UEFI before Nvidia's driver initializes.
No VBIOS flash and no driver patch, and the unlock disappears after a complete power-off.
The idea underneath came from Zagitta's earlier GPU Unlock project, which showed RTX 30 series cards reaching nearly twice their normal Tensor Core throughput on certain FP16 and BF16 workloads, with real AI image generation gaining close to 60 percent.
That version needed a Linux kernel module.
This one runs on Windows and Linux, covers RTX 20, 30 and 40 series, with GTX 16 still experimental and RTX 50 unsupported.
The early numbers: the 2070 climbed to roughly 60 to 63 TFLOPS, a 67 to 75 percent gain.
An RTX 4060 laptop went from 29.89 to 49.33 TFLOPS in FP16.
For anyone paying per hour, that shifts the arithmetic.
A rented RTX 3090 at $0.15/hr that does more matrix math in the same block of time means fewer hours bought for one fine-tune.
Buy Now Pay Later keeps the first month at zero upfront.
If your last training run had gained 70 percent more throughput per card, what would you spend the freed hours on?
https://ai.bhkcloud.com/?utm_source=linkedin&utm_medium=social&utm_campaign=bhk_social_2026w44&utm_content=evening
Originally posted on linkedin_personal