NVIDIA's older GeForce RTX graphics cards are gaining renewed attention after an open source tool demonstrated substantial improvements in artificial intelligence computing performance.
The utility, known as Zagitta GPU-Unlock, reportedly removes certain compute throughput restrictions on supported GPUs, allowing them to process FP16 and BF16 workloads considerably faster.
Initial testing shows improvements approaching 80% on several RTX 30 series models, while some RTX 40 series cards are also reportedly benefiting.
However, growing interest in using older gaming GPUs for AI workloads is coinciding with sharp increases in secondhand graphics card prices.
GPU Unlock Tool Improves AI Compute Performance Without Hardware Modifications
Zagitta uses a Linux kernel module to modify GPU configuration settings before NVIDIA's standard graphics driver initializes.
Unlike traditional modification techniques, the tool does not require flashing a modified BIOS or physically altering the graphics card.
It also avoids directly patching NVIDIA drivers, potentially reducing maintenance requirements following driver updates.
The approach primarily targets Ampere GPUs, including models based on GA102, GA104, and GA106 chips.
Reported benchmark improvements include:
| GPU | Reported BF16 compute improvement |
|---|---|
| RTX 3090 Ti | 68% to 80% |
| RTX 3090 | 25% to 67% |
| RTX 3080 10GB | 70% to 73% |
| RTX 3070 | 73% to 81% |
| RTX 3060 12GB | Approximately 80% |
Results vary depending on the GPU, workload, and configuration.
RTX 3090 Ti Shows Significant Improvements in AI Benchmarks
Testing on the RTX 3090 Ti reportedly increased BF16 throughput from approximately 87.5 TFLOPS to 157.4 TFLOPS.
TF32 performance also improved from 42.6 TFLOPS to 82.6 TFLOPS.
For large language model workloads, testing showed improvements of approximately 33% to 36% in prompt processing performance.
Time to first token decreased by around 25%, although token generation performance remained largely unchanged.
The utility currently supports Linux, with no official Windows version available.
Used RTX 30 Graphics Card Prices Reportedly Increase
Demand for older NVIDIA GPUs has reportedly increased as developers explore more affordable hardware for running AI models locally.
According to marketplace observations, RTX 3090 cards that previously sold for approximately $1,000 are now appearing above $2,000.
Some RTX 3090 Ti listings have reached $2,650, while RTX 3080 models are reportedly listed between $900 and $1,100.
These are observed asking prices rather than confirmed market averages.
Although the unlock makes older GPUs more attractive for AI computing, its performance benefits do not automatically extend to gaming workloads.
For gamers considering secondhand RTX cards, increased AI demand could make affordable upgrades harder to find.
Discussion (0)
Be the first to comment.