VideoCardz reports that separate modding projects have raised measured throughput for certain Tensor Core operations on selected GeForce RTX 20-, 30- and 40-series graphics cards. Its project list names Zagitta’s GPU-Unlock for RTX 30 cards, GPU-Unlock-Ada for RTX 40 cards and Turing Unlock for RTX 20 cards.
The reported results depend on the card and the test. VideoCardz cites an RTX 4090 test in which FP16 Tensor Core instruction throughput rose from 165.6 to 331.6 TFLOPS. It also reports an RTX 2080 Ti result of 111 TFLOPS, up from 59.9 TFLOPS before the unlock. That 2080 Ti had an unrelated physical modification giving it 22GB of VRAM.
For the RTX 30 project, Wccftech cites benchmarks published in the GPU-Unlock repository: At a 450W power limit, an RTX 3090 Ti rose from 87.50 to 157.41 TFLOPS in a dense BF16 matrix-multiplication test and from 42.58 to 82.61 TFLOPS in a TF32 test. Those benchmarks measure a different kind of work from the RTX 4090 instruction-throughput test.
Application results offer another perspective. In a Reddit post embedded by Wccftech, user under-_-null reports ComfyUI throughput gains on one RTX 3090 Ti of 58.6% with MOSS-Image SFT BF16, 57.2% with Qwen-Image 2.1 Turbo BF16 and 56.1% with Z-Image Turbo BF16. MOSS-Image SFT INT8 improved just 0.3% in the same user's tests. The user says the card had a 450W ceiling in both conditions and that each image configuration had 15 measured runs per condition after warmups.
VideoCardz says the tools change low-level GPU configuration registers before the NVIDIA driver loads; the changes clear when the GPU loses power. Most projects target Linux, while Windows implementations are experimental. Support has not been validated across every GPU variant, and unsupported configurations may fail to boot. Wccftech also cites a report of a successful unlock without the expected BF16 gain. It says there is no convincing evidence that the unlock improves gaming performance or DLSS execution times, so the compute figures should not be read as gaming-FPS gains.
Comments