Modders have discovered a way to unlock previously restricted AI processing performance on Nvidia’s GeForce RTX 20, 30, and 40-series graphics cards.
The modification removes a performance limit that Nvidia placed on certain AI calculations, allowing supported GPUs to complete some operations up to twice as fast.
According to VideoCardz, the RTX 4090 can achieve double its previous peak performance in specific AI tests, while the RTX 3090 Ti has shown improvements of up to 80% in certain calculations.
However, the improvements mainly apply to AI-related tasks rather than gaming.
How Modders Unlocked the Performance
Nvidia’s GeForce RTX graphics cards contain specialized hardware called Tensor Cores, which helps process artificial intelligence workloads.
However, Nvidia has limited the speed of certain calculations on its consumer graphics cards. Specifically, calculations that use FP32 accumulation run at half the speed of similar operations using FP16 accumulation.
These restrictions do not apply in the same way to some of Nvidia’s professional graphics cards. Modders have now discovered a way to bypass the restrictions by changing low-level GPU settings before Nvidia’s graphics driver starts.
The modification does not require replacing the GPU hardware or permanently changing its firmware. The changes are temporary and are cleared when the GPU loses power.
RTX 4090 and RTX 3090 Ti See Major Improvements
Early testing shows significant improvements in certain AI calculations.
The RTX 4090’s peak processing performance increased from approximately 165.6 trillion operations per second to 331.6 trillion operations per second in a specific Tensor Core test.
Meanwhile, the RTX 3090 Ti recorded an improvement of approximately 80% in BF16 calculations used for AI processing.
| Graphics Card | Performance Before | Performance After | Improvement |
|---|---|---|---|
| RTX 4090 | 165.6 TFLOPS | 331.6 TFLOPS | Around 100% |
| RTX 3090 Ti | 87.5 TFLOPS | 157.4 TFLOPS | Around 80% |
These figures represent specific AI computing tests rather than overall GPU performance.
Other supported graphics cards have also shown improvements, although the results vary depending on the model and workload.
What This Means for AI and Gaming
The unlocked performance could benefit people who use Nvidia graphics cards to run AI applications on their computers.
Tasks involving complex mathematical calculations, including AI image generation and processing prompts for local chatbots, could become faster.
However, the performance improvements are generally smaller in real-world applications than in the benchmark results.
For example, some RTX 3090 tests showed improvements of roughly 14% to 36% in AI prompt processing.
The speed at which chatbots generate text remains largely unchanged because that process often depends more on memory bandwidth than raw processing power.
Importantly, the modification does not provide a similar performance boost in video games.
Not All Nvidia Graphics Cards Are Supported
The modifications currently target selected models from Nvidia’s RTX 20, 30, and 40-series graphics cards.
Most of the available tools are designed for Linux, although experimental Windows versions are also being developed.
Support varies between different GPU models, and the modifications have not been confirmed to work across every graphics card in these generations.
Nvidia’s newer RTX 50-series cards are not currently covered by the confirmed working methods.
The tools are also experimental and carry risks, including system crashes, startup failures, and possible hardware problems.
For now, the discovery shows that some older Nvidia graphics cards can deliver significantly higher AI processing performance without requiring new hardware, although the actual benefits depend on the software and tasks being used.
The post Modders Find a Way to Make Nvidia’s Older GPUs 2x Faster appeared first on ProPakistani.