What we know

Modders have unlocked hidden AI performance on Nvidia GeForce RTX 20-, 30-, and 40-series GPUs by altering driver or firmware settings. This modification targets the tensor cores, which normally process FP16 or BF16 multiply-and-accumulate operations at half the rate of FP32 instructions on these consumer cards. By bypassing this limitation, some users have nearly doubled AI-related benchmark scores on affected GPUs.

Why it matters

Nvidia’s GeForce RTX GPUs rely on tensor cores for AI and machine learning workloads, but these cores are intentionally limited in consumer models compared to professional-grade hardware. Unlocking the full throughput of FP16/BF16 operations could significantly boost AI performance for gamers, researchers, and developers using these GPUs. This discovery suggests that current hardware may have untapped potential, which could impact how AI workloads are run on widely available consumer graphics cards.

What is still unknown

No independent tests have verified the accuracy or stability of the unlocked performance. It is unclear which AI frameworks or languages fully support these modifications. There is no information on whether Nvidia will address this through official updates or patches, nor is there a timeline for any broader availability of this unlocked mode.