Modder Unlocks AI Tensor Core Limits On Older NVIDIA RTX 20, 30, And 40 Series GPUs

Modders have figured out a way to unlock the Tensor Core compute limits on older NVIDIA RTX 20, 30, and 40 Series graphics cards. With the limiters removed, they say that the cards will be able to run FP16, BG16, and TF32 math with FP32 accumulation at full speed.

One modder, who posted the mod on GitHub, lists down the supported GPUs, but it should be made clear that at current, the gpu-unlock mode currently only runs stable with RTX 30 Series GPUs. As for the RTX 20 and 40 Series, it’s still experimental.

“NVIDIA ships GeForce cards with their tensor cores capped to half speed for FP16, BF16 and TF32 math with FP32 accumulation, which is what nearly all AI training and unquantized inference uses. The same silicon sold as an RTX A-series or professional card runs at full rate. gpu_unlock removes that cap at boot, before the stock NVIDIA driver loads. No patched driver, no VBIOS flash, nothing to redo when NVIDIA releases a new driver.”

Interestingly, the gpu-unlock mod also works with laptop GPUs. In the case of the RTX 30 series, that list seems limited to RTX 3070 and 3080 laptop GPUs. Also, to be clear, the unlocks do not add memory bandwidth, CUDA Cores, or even increase gaming performance. The mod seems strictly for AI and compute work gains only.

Upon unlocking the Tensor cores, the modders say that AI performance on RTX 30 and 40 Series cards nearly doubled. The list the modders gave are as below:

Again, this unlock isn’t for gaming, but more for the benefit of AI computing. Also, the modder doesn’t believe that this mod will damage the hardware, pointing out that the safeguards, such as throttling and power management stay in place.

(Source: Techspot, TweakTown, GitHub)