Open-source project unlocks extra performance on Nvidia RTX 3090

Unlocking untapped potential

Nvidia’s six-year-old GeForce RTX 3090 may have more computing power than it currently uses in standard configurations. A new open-source project called gpu-unlock, created by GitHub user Zagitta, aims to reveal this hidden potential. The project focuses on improving performance for specific types of AI-related math known as Tensor Core workloads.

According to the developer, the current hardware is limited by an instruction-issue rate that affects certain operations. Specifically, these limitations impact FP16 and BF16 Tensor Core operations when they use FP32 accumulation. These are common formats used in deep learning and high-performance computing tasks.

The project is notable because it does not require any physical hardware modifications or a patched Nvidia driver. It essentially changes how the hardware selects its speed for these specific operations. This allows the card to run at speeds closer to what is found on professional Nvidia GPUs.

How the unlock works

The technique used by Zagitta builds on previous research that was originally used to unlock capabilities on Nvidia CMP mining GPUs. It works by targeting the hardware speed-selection registers to remove the existing restrictions. By modifying these registers, the card can execute more instructions per cycle than it normally allows.

The implementation relies on a Linux kernel module that runs before the standard Nvidia graphics driver loads. This module modifies the relevant registers and then resets the GPU before handing control back to the unmodified driver. Because it works this way, the user does not have to permanently flash the card’s VBIOS or any other firmware.

This non-permanent approach makes the process safer for the hardware than traditional flashing methods. However, because it is an experimental software solution, it currently requires a Linux environment. A separate experimental port for Windows has been tested but is not yet a standard feature.

Measured performance gains

The results of this unlock vary significantly depending on the specific workload being performed. In one community report, an RTX 3090 saw its BF16 matrix multiplication performance jump from 77.5 TFLOPS to 129.6 TFLOPS. This represents a nearly two-fold increase in raw computing throughput for that specific operation.

Other tests showed different levels of improvement across various hardware models. The RTX 3090 Ti saw its BF16 throughput increase from 87.5 TFLOPS to 157.4 TFLOPS. Additionally, TF32 performance on an RTX 3090 rose from 40.2 TFLOPS to 69.8 TFLOPS during testing.

Real-world applications show mixed results depending on the type of AI task. One user reported a 16% improvement in processing long-context prompts for a local large language model. However, that same user found that the speed of generating new tokens remained almost completely unchanged.

Impact on creative workflows

For users involved in AI-assisted image generation, the gains may be more substantial. Testing with ComfyUI on an RTX 3090 Ti showed higher throughput across several BF16-heavy workloads. This suggests that the unlock is particularly effective for certain types of generative art tools.

There is also a possibility that Nvidia’s DLSS technology could see some benefits since it also utilizes Tensor Cores. DLSS is a technology that uses AI to upscale images and improve frame rates in video games. However, no independent benchmarks have confirmed any actual improvements in gaming performance yet.

While the potential for extra power is intriguing, the project remains in an experimental stage. Users who install it on their primary gaming PC may face risks since it is not an officially supported method. It currently serves as a way for adventurous users to squeeze more life out of older hardware.