Zagita tool doubles RTX 30/40 GPU compute, driving used prices up
A new open-source Linux tool called Zagita has unlocked the full BF16/FP16 compute rates on NVIDIA's RTX 30 and RTX 40 series GPUs. By using a kernel module that applies specific speed selectors before the driver binds, the tool doubles the instruction issuing speed for FP/BF16 FMA operations on GA102, GA103, GA104, and GA106 chips without requiring vBIOS flashing or hardware mods. Users report…
Key points
- Zagita tool doubles BF16/FP16 compute on RTX 30/40 GPUs via Linux kernel module.
- RTX 3090 Ti BF16 performance increases from 87.5 TFLOPs to 157 TFLOPs.
- Used RTX 3090 prices have surged past $2,000 due to increased AI demand.
In practical AI LLM workloads, the unlock delivers a 33-36% faster prefill speed and reduces time-to-first-token latency by 25%, though decode performance remains unchanged. One user reportedly achieved 381 tokens per second on a Qwen3.8-27B model using an RTX 3090, a significant increase from the baseline 82 TPS. The tool currently supports only Linux, with Windows support not yet available.
This performance uplift has triggered a sharp rise in the used market. The RTX 3090 24 GB, previously around $1,000, has climbed past $2,000, while the RTX 3090 Ti approaches $3,000. RTX 3080 models are now selling for $900-$1,100 on eBay. While the tool makes older cards viable for AI inference and fine-tuning, it is making them less accessible for gamers due to the price surge driven by AI demand.
The story so far
2 episodes →- Zagita tool doubles RTX 30/40 GPU compute, driving used prices upthis story
CMP170HX inspiration unlocks 30x/40x also, up to 2X Prefill speedup --> NVIDIA RTX 30 & RTX 40 GPUs Get Impressively Fast In AI LLMs As Compute Doubles With New Unlock Tool But Prices Surge Too
wccftech.com · 24 October 2026
Loading the full article…
This text was published by wccftech.com and written by Hassan Mujtaba. It is reproduced here with attribution so you can read it in full; the rights remain with the publisher. Read it at the source ↗
Coverage and discussion
1source- Reddit discussionreddit.com
The headline, key points and digest above were generated by Digest AI's editorial model from the linked sources. Automated summaries can contain errors: the sources are the record. Spotted a mistake? Tell us. Published by Martin K., who runs Digest AI and handles corrections.
More in Hardware & Compute
All →- Nvidia H200 rental costs rise about 160% since January, report shows · 1 src
- Nvidia AI revenue grew 117%, Broadcom grew 221% last quarter · 1 src
- Nvidia and SPIL plan 15x expansion of co-packaged optics capacity by mid-2027 · 1 src
- Investors watch AMD, Google, and Amazon challenge NVIDIA in AI chips · 2 src
- Google and Voltus expand virtual power plant deal for 1,070 megawatts in PJM grid · 1 src
Comments
via GitHub Discussions