NVIDIA RTX 30 & RTX 40 GPUs Get Impressively Fast In AI LLMs As Compute Doubles With New Unlock Tool But Prices Surge Too - Wccftech
A new open-source tool called Zagitta has unlocked the full AI compute potential of NVIDIA's RTX 30 and RTX 40 series GPUs, doubling their FP16/BF16 performance for AI workloads.
The open-source Zagitta tool is unlocking previously restricted AI compute capabilities in NVIDIA's RTX 30 and RTX 40 series GPUs, delivering performance boosts of up to 80% without requiring hardware modifications or vBIOS flashing. The tool works by modifying kernel timing before the NVIDIA driver binds to the card, effectively doubling the FP16/BF16 instruction issuing speed on GA102, GA103, GA104, and GA106 GPUs.
Performance Gains
Users report significant improvements across the RTX 30 series, with the RTX 3090 Ti seeing its BF16 capabilities jump from 87.5 TFLOPs to 157 TFLOPS, while TF32 performance increases from 42.6 TFLOPs to 82.6 TFLOPs. The RTX 3080 10GB similarly shows gains from 65 TFLOPS of FP16 to 113 TFLOPS at its stock 320W power consumption.
3090 Ti dense 8192³ GEMM at 450 W, five-sample medians with FP32 accumulation
AI Workload Improvements
In AI language model applications, the unlocked RTX 30 GPUs deliver 33-36% faster Prefill output for long-context processing while reducing Time-to-first-token latency by 25%. Matrix-multiply performance remains unchanged, and decode performance sees no impact. The tool also extends to RTX 40 series GPUs, potentially increasing their second-hand market value.
I believe this might end up increasing these GPUs' 2nd-hand prices as well. It'll also be interesting to see if this can be leveraged to improve DLSS perf since it runs on Tensor Cores.
Market Impact
The performance improvements are driving significant price increases in the used GPU market. The RTX 3090 24GB has climbed from around $1000 to over $2000, while the RTX 3090 Ti is approaching $3000. RTX 3080 10-12GB models now sell for $900-$1100 on eBay. This surge reflects growing demand from AI enthusiasts and researchers seeking cost-effective solutions for large language model fine-tuning and inference.
The price increases come alongside reports of optimized AI models running on older hardware. A user reportedly achieved 381 tokens per second on Qwen3.8-27B using a single RTX 3090, a 4.6x improvement from the baseline 82 tokens per second through optimized MTP, DFlash2 with context lookup, and verification techniques.
Qwen3.8-27B reportedly reached 381 tokens/sec on a single RTX 3090 with 24GB of VRAM. The starting point was 82 tokens/sec. That is approximately a 4.6× increase without changing the GPU or the model.
Platform Limitations
Currently, the Zagitta unlock tool is only available on Linux platforms, with Windows support not yet implemented. While the RTX 30 series sees the most significant benefits, the tool also extends to older GTX 9, GTX 10, and RTX 20 series cards, potentially extending their viability for AI workloads.
Related
