- Community projects have demonstrated roughly 2× peak FP16/BF16 tensor instruction throughput on selected RTX 4090 cards and substantial gains on some RTX 3090-series cards. <sup>[1]</sup>
- The unlock requires control of the host and early-boot code; the documented Linux method uses a kernel module before Nvidia’s driver binds, while experimental Windows work uses pre-boot firmware. <sup>[2]</sup>
- Nvidia had not publicly commented or announced a driver or firmware fix as of October 26, 2026. <sup>[3]</sup>
Researchers and modders have shown that selected Nvidia GeForce RTX 20-, 30- and 40-series GPUs can run certain AI-oriented Tensor Core operations faster than their factory settings allow, using community-built low-level software rather than a hardware modification. [1]
The work affects specific GPU architectures and models, rather than every card in those product families. The main Ampere project targets selected RTX 30-series cards, while separate projects cover Turing RTX 20-series and Ada RTX 40-series hardware. The project documentation warns that some cards complete the unlock without producing a measurable throughput gain. [2]
A local, unsupported modification #
The Linux implementation loads before Nvidia’s stock kernel driver, changes GPU speed-selection registers and resets the device before handing control to the unmodified driver. The documented method uses Nvidia firmware inputs and does not require flashing a VBIOS or patching the Nvidia driver. [2]
That makes this a local-control issue rather than a conventional remote vulnerability: an attacker would already need administrative control of the machine, physical access, or an equivalent ability to control its boot process. The sources reviewed do not show a remote attack path or a bypass of Nvidia’s authenticated production firmware. [2]
Performance gains are workload-specific #
Project measurements report approximately 2× FP16- and BF16-with-FP32-accumulation instruction throughput on a tested RTX 4090, while an RTX 3090 Ti benchmark showed BF16 throughput rising from 87.50 to 157.41 TFLOPS. <sup>[1]</sup> Those are low-level or narrowly selected compute results, not a guarantee of equivalent gains in model training, inference or gaming.
Independent coverage found that the original project was limited to Linux and that no independently verified gaming performance improvement had been established. Nvidia had not issued a public response or patch when the reports were published. [3]
Companies mentioned #
Further sources #
The stories that matter, in one email. Free — unsubscribe anytime.