Modders bypass Nvidia’s AI performance limits on RTX 20, 30, 40-series GPUs
Modders have found a firmware exploit to remove Nvidia’s imposed half-speed AI performance limits on RTX 20, 30, and 40-series GPUs, according to VideoCardz. The unlock doubles the RTX 4090’s peak tensor rate and improves matrix math by up to 80% for the RTX 3090 Ti, though real-world gains vary by workload and card. Nvidia’s limit restricts FP32-accumulate math to half the FP16-accumulate rate,…
Key points
- Modders bypass Nvidia’s AI throttle via firmware exploit, doubling RTX 4090 tensor rates and boosting RTX 3090 Ti matrix math by up to 80%
- Three projects offer unlocks: **gpu-unlock** (Ampere), **gpu-unlock-ada** (Ada, partial), and **turing-unlock** (Turing), with varying card support
- Risks include hardware damage or system instability; gains vary by workload and power limits, with LLM decode unchanged
The exploit’s risks include potential hardware damage, system hangs, or data corruption, per project warnings. Performance improvements depend on power limits—cards capped at 225W saw minimal gains (e.g., +25% BF16 on RTX 3090), while higher-power variants achieved up to +85% in instruction rate tests. LLM prefill and diffusion tasks sped up 14–36% on Ampere, but LLM decode and INT8 performance remained unchanged. Hopper and Blackwell cards are not supported, and Ada/Turing coverage remains incomplete, with maintainers urging owners to report compatibility data.
The story so far
3 episodes →- Modders bypass Nvidia’s AI performance limits on RTX 20, 30, 40-series GPUsthis story
Nvidia RTX 20, 30, and 40-series GPUs hacked to remove a half-speed AI throttle
Tom's Hardware · 26 October 2026
Loading the full article…
This text was published by Tom's Hardware and written by Shane Downing. It is reproduced here with attribution so you can read it in full; the rights remain with the publisher. Read it at the source ↗
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 →- Meta launches $349 Ray-Ban Audio glasses with no camera and AI integration · 1 src
- Monolithic Power Systems and Analog Devices linked to AI data center power demand · 1 src
- Amazon Bee review: $75 AI wearable summarizes conversations and suggests actions · 2 src
- Leaked footage shows Chinese factories converting RTX 5090s for AI · 1 src
- Nvidia CEO Jensen Huang says quantum computers could help train AI · 1 src
Comments
via GitHub Discussions