NVIDIA GeForce RTX 30 GPUs Get Up To 2X Faster AI Compute Performance With New GPU Unlock Mod, Real-World Gains Of Up To 60% – USFTB
NVIDIA’s aging GeForce RTX 30 Series (codename Ampere) GPUs apparently have a considerable amount of untapped computing potential. A new open-source project called GPU Unlock, developed by GitHub user Zagitta, allows Ampere GPUs to overcome certain artificial performance restrictions, nearly doubling their Tensor Core compute throughput in specific workloads without requiring a modified NVIDIA driver or a permanent VBIOS flash.
The project specifically targets the instruction issue-rate limitation affecting FP16/BF16 Tensor Core operations with FP32 accumulation, which operates at a reduced rate on consumer Ampere graphics cards compared to their professional counterparts. By removing this restriction, the tool effectively enables additional computing performance that the GPUs were already capable of delivering.
Interestingly, GPU Unlock builds upon earlier research involving NVIDIA’s CMP cryptocurrency mining GPUs. It combines the firmware unlocking techniques of the CMP project with an early initialization approach, using a Linux kernel module that executes before NVIDIA’s GPU driver loads. The module modifies the relevant hardware speed selectors, performs a Function Level Reset (FLR), and then allows the unmodified driver to initialize the GPU normally.
This also means that users won’t have to patch their NVIDIA drivers every time a new version is released. The developer currently lists support for GA102, GA103, GA104, and GA106 GPUs, covering a substantial portion of the GeForce RTX 30 Series lineup, although compatibility and performance improvements aren’t guaranteed across every graphics card.
The performance improvements can be remarkable. According to benchmarks published in the project’s GitHub repository, an RTX 3090 Ti operating at a 450W power limit delivered the following results in dense matrix multiplication (GEMM) tests with FP32 accumulation:
| Precision | Stock | GPU Unlock | Improvement |
|---|---|---|---|
| FP16 | 87.46 TFLOPS | 118.39 TFLOPS | +35% |
| BF16 | 87.50 TFLOPS | 157.41 TFLOPS | +80% |
| TF32 | 42.58 TFLOPS | 82.61 TFLOPS | +94% |
These results aren’t exclusive to the RTX 3090 Ti, either. An RTX 3080 Ti owner reported BF16 throughput increasing from 71.59 to 113.98 TFLOPS, while TF32 performance improved from 35.65 to 65.28 TFLOPS, with the GPU limited to 315W. Another RTX 3080 10GB compatibility report documented FP16/BF16 performance increasing from approximately 65 TFLOPS to 110–113 TFLOPS under the card’s original 320W power limit. These improvements were also highlighted on X.
But what about actual application performance? Fortunately, community members have already started testing the unlock in real-world AI workloads, and the preliminary findings are encouraging.
One RTX 3090 Ti user on Reddit benchmarked GPU Unlock in ComfyUI, reporting 58.6% higher throughput with MOSS-Image SFT BF16, 57.2% with Qwen-Image 2.1 Turbo BF16, and 56.1% with Z-Image Turbo BF16. Meanwhile, INT8 configurations saw much smaller improvements, with MOSS-Image SFT INT8 improving by just 0.3%. The tester also reported identical decoded image output between the stock and unlocked configurations.
Large Language Model (LLM) inference is another potential beneficiary, although improvements are considerably more workload-dependent. For example, an RTX 3090 compatibility report documented a 33% improvement in vLLM prompt-processing/inference throughput, while token generation remained practically unchanged. That’s understandable, considering that LLM decoding is frequently limited by memory bandwidth rather than raw Tensor Core compute throughput.
An additional development worth highlighting is an experimental Windows implementation, which uses a UEFI-based approach instead of the original Linux kernel module. Its developer has documented a successful RTX 3090 implementation, with BF16 compute throughput increasing by approximately 82.5%. However, this remains an experimental, hardware-specific solution rather than a generally compatible Windows release.
There are also important limitations. GPU Unlock doesn’t increase memory bandwidth, enable additional CUDA cores, or automatically improve every workload. At least one RTX 3080 Ti 20GB compatibility report documented a successful unlock procedure without the expected BF16 performance improvement, demonstrating that results aren’t universal.
As for gaming, NVIDIA’s DLSS technologies also utilize Tensor Cores, but there’s currently no convincing evidence that GPU Unlock improves gaming performance or DLSS execution times. The affected instructions and accumulation modes need to be used by the actual workload for any benefit to materialize. Consequently, the impressive compute benchmarks shouldn’t be mistaken for an equivalent increase in gaming FPS.
Nevertheless, GPU Unlock is an intriguing development for GeForce RTX 30 Series owners, particularly those using their GPUs for local AI inference, image generation, and other compute-intensive applications. While the project is still experimental, the prospect of extracting substantially more performance from existing hardware without permanently modifying it is certainly appealing.
News Source: GPU Unlock GitHub repo
Follow Wccftech on Google to get more of our news coverage in your feeds.
