An open-source USB-bootable tool can lift FP16 performance on some Nvidia GeForce RTX 20-, 30- and 40-series cards by removing a tensor-performance limit. Tom’s Hardware reports gains of up to 65% in the developer’s tests, with important compatibility and setup caveats for anyone tempted to give an older GPU a second job in AI computing.
Watch Desk analysis
What happened
The project, called nv-uefi-unlocker, runs before the Nvidia driver loads and changes firmware settings without flashing the card’s VBIOS or modifying its driver. Its reported Windows results include a 65% FP16 gain on the developer’s RTX 4060 laptop GPU. Other reported results vary by card and test, and should not be compared directly across models.
The report says RTX 20-, 30- and 40-series cards are supported to varying degrees, but support does not mean every model has been tested. RTX 50-series cards are unsupported. The tool needs a UEFI boot, and Secure Boot must be disabled or configured to trust a self-signed EFI. The unlock also has to be run again after power loss.
Why it matters
This is a concrete way to repurpose some existing graphics cards for more AI compute, rather than buying new hardware. But the headline percentage is not a blanket promise: results depend on the card, workload, power and thermals, and some models have known issues or remain untested.
The setup also changes the boot process. The report notes that a per-card ROM hash check is not yet built in, and warns that combining the unlock with a power-limit modification can exceed factory power and thermal limits. This is a hands-on experiment, not a tidy software switch.
Our read
The appealing part is squeezing more tensor performance from hardware people already own. The less appealing part is that the route there involves firmware, boot settings and a card-by-card compatibility picture. For AI tinkerers with a supported GPU, the reported gains are worth investigating; for everyone else, wait for clearer testing and safeguards before treating this as a casual upgrade.
What to watch
- Whether the project adds the planned ROM hash check.
- Which additional RTX 20- and 40-series models are tested, and what problems emerge.
- Whether results hold across more AI workloads and sustained thermal conditions.
Discussion spark: Would you trade Secure Boot convenience and a more involved boot process for extra AI performance from a GPU you already own?
Sources and evidence
Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.