By Anil Ganti
Publication Date: 2026-10-24 04:01:00
Nvidia’s six-year-old GeForce RTX 3090 apparently has some untapped computing potential. A new open-source project called gpu-unlock, developed by GitHub user Zagitta, can reportedly deliver double-digit improvements in certain Tensor Core workloads without requiring hardware modifications or a patched Nvidia driver.
The project targets an instruction-issue rate limitation affecting FP16 and BF16 Tensor Core operations with FP32 accumulation on Ampere-based GeForce cards. According to the developer, the affected operations can run at roughly half the rate available on professional Nvidia GPUs. The unlock changes hardware speed-selection registers to remove that restriction.
Interestingly, the technique builds upon earlier research used to unlock capabilities on Nvidia CMP mining GPUs. Zagitta’s implementation uses a Linux kernel module that executes before Nvidia’s graphics driver loads, modifies the relevant registers and resets the GPU before handing control back to the unmodified driver. As a result, it does not require permanently flashing the card’s VBIOS or modified firmware.
In one community compatibility report, an RTX 3090 achieved 129.6 TFLOPS in BF16 matrix multiplication, up from 77.5 TFLOPS. TF32 performance increased from 40.2 to 69.8 TFLOPS. Its sibling, the RTX 3090 Ti, saw BF16 throughput jump from 87.5 to 157.4 TFLOPS.
That said, the gains are very workload-dependent. Another RTX 3090 user reported a 16% improvement in long-context prompt…
