A hidden speed switch inside gaming graphics cards

What happened

The graphics card in a high-end gaming PC, like NVIDIA's RTX 4090, is also what a lot of people use to run AI at home or in a small office. A researcher on GitHub, who goes by JustAResearcher, published a project that rewrites a small piece of the card's built-in software, its firmware, so it handles one specific kind of AI math faster. It appeared on October 23, 2026, and within a week it was on GitHub's trending list with 4,453 stars.

In one test, that math ran about twice as fast. TechSpot reported the rate jumping from 165.6 to 331.6 TFLOPS, and other developers have already adapted the idea for older RTX 20 and 30 series cards.

The honest version

A DEV Community write-up adds the part worth remembering. The doubling came from a narrow benchmark. On the kind of work AI models actually do, gains were 12–34 percent, and each figure comes from a single pair of runs. You build the firmware yourself on a Linux computer you control, and cards other than the 4090 are marked experimental. This is a risky hardware hack, not a free upgrade.

Why it's still interesting

It's a quiet reminder that the machines you already own may have more in them than the spec sheet admits. If your business runs AI on its own hardware, this is worth watching from a comfortable distance while others test it.

Words worth knowing

• GPU: the graphics card. Built for games, it turns out to be very good at AI math.
• Firmware: the software baked into a device that tells it how to behave.
• Benchmark: a standard test used to measure speed.

If someone pitches you this as free speed, smile and ask for the real-world numbers.