Zagitta’s Nvidia RTX 30 GPU unlock mod is pushing GeForce RTX 30-series cards past stock Tensor Core output in compute benchmarks. On a GeForce RTX 3090, BF16 throughput rises from 77.5 TFLOPS to 129.6 TFLOPS, which puts the tool squarely in the path of owners who run local AI and other compute-heavy workloads.
Zagitta and gpu-unlock
Notebookcheck reported on October 24, 2026 that gpu-unlock can lift results by up to 80% in certain compute benchmarks on NVIDIA’s RTX 30-series cards. The project is open source, and it is hosted on GitHub under Zagitta’s user name, which makes the tool easy to inspect but also easy to test against its own claims.
GeForce RTX 3090 BF16
The reported gain comes from a Linux-based kernel module that loads before the official NVIDIA graphics driver. That load order is the trick, because the tool changes how fast one class of instructions can issue instead of pushing voltage or clocks beyond factory limits.
Notebookcheck says the target is an instruction-issue limit that affects certain Tensor Core operations on Ampere-based GeForce cards. Those operations are FP16 and BF16 math with FP32 accumulation, so the practical change is not a broad system tweak but a narrower shift in how quickly those compute paths can run.
RTX 30-series compute limits
The developer says GeForce parts run these instructions at roughly half the rate that NVIDIA’s professional cards manage. That gap explains why the unlock matters to owners of NVIDIA cards who use them for AI and similar workloads, because the tool is aimed at compute throughput rather than raw gaming frame rates.
The complication is that the project still carries an experimental label, and the source says nobody has shown that the unlock helps games. So the reader-facing decision is blunt: if the card is used for BF16, FP16, or FP32 accumulation workloads, the mod may be worth watching; if the only goal is gaming, the reported gains do not yet change the case.
What remains unresolved is gaming performance, because the available result set stops at compute benchmarks. Until a broader compatibility report appears, RTX 30-series owners are left with one hard number — 129.6 TFLOPS on a GeForce RTX 3090 — and one open question about whether that higher ceiling translates beyond AI and compute work.
