NVIDIA RTX 20, 30 and 40 GPUs Get AI Performance Unlock

 

Modders have developed experimental tools that remove an NVIDIA-imposed Tensor Core performance restriction on selected GeForce RTX 20, 30, and 40-series graphics cards, delivering substantial gains in certain AI computing workloads.

The restriction affects matrix operations that use FP16 or BF16 inputs with FP32 accumulation. On supported consumer GPUs, these operations normally run at a reduced instruction rate compared with professional NVIDIA cards.

Developer Zagitta's open-source gpu-unlock project on GitHub targets RTX 30-series Ampere GPUs, while JustAResearcher's gpu-unlock-ada extends the approach to selected RTX 40-series hardware.

Separate Turing research also demonstrates the technique on RTX 2080 Ti cards, although support varies significantly between models.

On an RTX 4090, the Ada project reports FP16 Tensor Core throughput increasing from 165.6 to 331.6 TFLOPS. However, actual matrix multiplication benchmarks show smaller improvements, generally between 12% and 34% for FP16.

An RTX 3090 Ti achieved approximately 80% higher BF16 matrix multiplication performance in one reported test, while other Ampere cards showed different results depending on workload and power limits.

The technique changes low-level GPU configuration before the NVIDIA driver initializes. It does not require a permanent VBIOS modification, and the changes disappear after a full power cycle.

These gains primarily benefit certain AI calculations, including model processing and image generation. They do not establish higher gaming frame rates, faster DLSS performance, or doubled application speed.

RTX 50-series GPUs are not supported by these projects.

The modifications remain unofficial and experimental, with developers warning about possible crashes, data corruption, and hardware damage.