Modders unlock hidden AI performance on older NVIDIA GeForce GPUs

Bypassing hardware limits

Modders have discovered a method to remove performance restrictions on several generations of NVIDIA GeForce graphics cards. This discovery began with the RTX 3090 series but has since expanded to include RTX 40 models and even the older RTX 2080 Ti. These modifications allow for a significant boost in AI-specific processing power without requiring any physical hardware changes.

The process does not involve flashing a new VBIOS or modifying the standard graphics drivers. Instead, the tools work by altering low-level GPU configuration registers before the NVIDIA driver actually loads. This allows the hardware to issue certain Tensor Core instructions at a much higher rate than NVIDIA originally intended.

Because these changes occur at the register level, the modifications are temporary in nature. The settings are cleared automatically every time the GPU loses power. This means users do not have to worry about permanent changes to their hardware configuration or driver stability.

Quantifiable performance gains

The results of these unlocks are substantial when measured in specific compute tests. For example, an RTX 3090 Ti saw its BF16 matrix multiplication performance jump from 87.5 to 157.4 TFLOPS. Similarly, an RTX 4090 saw its measured FP16 Tensor Core instruction throughput double from 165.6 to 331.6 TFLOPS.

Other cards have shown similar improvements, including an RTX 2080 Ti that reached 111 TFLOPS compared to its original 59.9 TFLOPS. While that specific card had been physically modified with more memory, the software unlock provided the additional compute boost. Successful tests have also been confirmed on the RTX 3060, 3070, and 3080 models.

These gains primarily affect FP16 and BF16 operations with FP32 accumulation. They also impact certain TF32 and FP8 workloads on the supported architectures. These specific types of math are the backbone of many modern artificial intelligence applications.

Impact on AI and gaming

While the numbers look impressive, these unlocks do not provide any extra frames per second in video games. Gaming performance remains unaffected because the modifications are targeted specifically at tensor operations. The primary beneficiaries of this unlock are users running AI tasks like large language model (LLM) prompt processing and image generation.

In real-world testing, an RTX 3090 showed improvements of roughly 14% to 36% when processing LLM prompts. However, token generation speeds remained largely the same for many users. This is because token generation is often limited by memory bandwidth rather than the raw compute throughput that these modders have unlocked.

The tools currently target Linux systems as the primary platform for these modifications. While experimental versions for Windows are available, they are not yet fully validated. Users should be cautious, as unsupported configurations have the potential to cause the system to fail to boot.

Market implications for users

The existence of these unlocks has sparked a debate about whether NVIDIA should officially remove these restrictions. Many of these cards are no longer in production and are out of warranty, making them prime candidates for performance boosts. Owners who use their existing hardware for AI work could see a significant boost in productivity without spending more money.

However, there is a potential downside for the general gaming community. Cards like the RTX 3090 and 4090 are already popular with AI researchers due to their high memory capacity. If their compute performance is also unlocked, they may become even more desirable on the second-hand market.

This increased demand from the AI sector could potentially drive up prices for used hardware. This creates a difficult situation where making older cards more capable for AI enthusiasts might make it harder for average gamers to find affordable used equipment. As VideoCardz.com reported, the GPU market is already in a delicate state, making these shifts noteworthy.