Open Source Unlocks RTX 30 Series BF16 Power, 3090 Ti Performance Nearly Doubles

Open-Source Project Unlocks Enhanced AI Performance on Nvidia Consumer GPUs

A new open-source project, Zagitta/gpu-unlock, is enabling enhanced AI computation capabilities on Nvidia’s Ampere-generation consumer graphics cards by unlocking their FP16/BF16 compute performance. The project utilizes a Linux kernel module to bypass driver limitations, allowing cards like the RTX 3090 Ti to achieve significantly higher throughput. Initial tests on the RTX 3090 Ti have shown BF16 performance soaring from 87.50 TFLOPS to 157.41 TFLOPS. Community reports suggest that large language model (LLM) prefill operations can see speedups of 10% to over 30%, leading to a reported surge in the second-hand market value of these GPUs, potentially making them less accessible for pure gaming enthusiasts.

Consumer GPUs Targeted by AI Compute Market

Consumer-grade graphics cards are once again drawing the attention of the artificial intelligence compute market. The open-source project Zagitta/gpu-unlock employs a Linux kernel module to inject speed selectors before Nvidia’s driver bindings take effect. This effectively brings the FP16/BF16 FMA launch rates of Ampere-generation GA102, GA103, GA104, and GA106 consumer cards closer to professional-grade performance. Crucially, this bypass does not require flashing the vBIOS or applying driver patches with every update.

According to the project’s README, dense GEMM benchmarks on an RTX 3090 Ti, operating under a 450W power limit, demonstrated BF16 performance increasing from 87.50 TFLOPS to 157.41 TFLOPS. TF32 performance rose from 42.58 TFLOPS to 82.61 TFLOPS, and FP16 saw an increase from 87.46 TFLOPS to 118.39 TFLOPS. Following a report by Wccftech on October 24th, discussions regarding these performance gains have intensified in the second-hand market.

Shift in Demand from Mining to AI

During the cryptocurrency mining boom, consumer cards experienced locked compute capabilities, only to see their value plummet after Ethereum’s transition. Now, with the resurgence of demand for local large model inference and fine-tuning, the same Ampere hardware is being repriced. The buyers have shifted from crypto miners to AI enthusiasts, and the supply-demand dynamics have moved from “coin mining” to “running models.”

The project’s documentation explains that this module combines CMP unlock techniques with pre-driver timing methods, achieving the unlock before Nvidia’s driver bindings are applied. The advantages include eliminating the need to flash the vBIOS and avoiding the recurring task of patching drivers after each update, thereby reducing long-term maintenance costs.

Current Status and Risks

The primary platform for this project is currently Linux, with Windows not yet offering official support for the kmod. The community also cautions about experimental risks, such as the need to disable settings like Secure Boot. The impact on stability and warranty can vary by hardware, and it is not recommended for use in formal production environments.

Reported compatible models include the RTX 3090 Ti, 3090, 3080, 3070, and 3060 12GB, among others. Community cases have also documented successful implementations with multi-GPU setups and virtual machine passthrough. However, not all attempts have been successful, with a few OEM cards reportedly remaining at half-speed even after the process.

This publication is providing an overview of the technical aspects and community feedback and does not offer installation instructions. Individuals interested in attempting this modification should independently assess the hardware and cybersecurity risks and refer to the project’s official documentation.

Performance Gains and Considerations

The compatibility reports compiled in the README indicate that BF16 GEMM gains typically range from 25% to 80%, with power limits playing a significant role. Cards with restricted power often see considerably smaller improvements, underscoring the importance of not solely focusing on peak performance figures.

For large language models (LLMs), prefill operations (long context pre-filling) have shown approximately 14% to 36% acceleration, leading to reduced first-token latency (TTFT). Workloads that are matrix-intensive, such as fine-tuning and operations like FlashAttention, tend to benefit more.

Conversely, workloads bottlenecked by decode, INT8, or memory bandwidth often experience minimal to no improvement. GitHub issues have also reported near-zero difference for llama.cpp Q8, indicating that quantized inference may not necessarily benefit.

In essence, the unlock modifies instruction launch rates and is not a universal solution. Before purchasing a graphics card, users should clarify whether their bottleneck is compute power or bandwidth to avoid misinterpreting “theoretical doubling” as a direct real-world performance increase.

Community posts have also suggested that Ada Lovelace (RTX 40 series) Tensor throughput could nearly double. However, the official README primarily focuses on the 30 series Ampere cards, and claims regarding the 40 series have not been fully verified and should be approached with caution.

Additionally, a rumor about a single RTX 3090 achieving 381 tokens/s for Qwen has circulated. This optimization is attributed to model-side techniques like MTP and DFlash and should not be conflated with the current unlock, to prevent overestimating the marginal benefits of hardware unlocking.

Second-Hand Market Surge and Implications

According to Wccftech’s compilation and community discussions, eBay second-hand prices have recently surged. RTX 3090 cards have risen from approximately $1,000 to over $2,000, the 3090 Ti is approaching $3,000, and the 10-12GB 3080 models are priced between $900 and $1,100. If the resurgence in second-hand demand is indeed driven by AI applications, pure gaming buyers may find it increasingly difficult to find bargains, and the price differential between new and old generation graphics cards could be reshuffled.

As cloud computing costs remain high and local LLM demand grows, older generation consumer cards that can achieve near-professional-grade half-precision throughput after unlocking are being rapidly repriced. This continues the trend of ex-mining cards finding new utility, but this time with AI players as the buyers. It further blurs the lines between consumer GPUs and artificial intelligence hardware, a development worth continued observation.