Tensor Core Unlock: New Mod Doubles Compute Power for GeForce RTX 20, 30, and 40
Modders have developed a way to bypass the artificial software limits placed on consumer-oriented NVIDIA GeForce RTX GPUs. The new RTX Unlock tool targets the RTX 20, RTX 30, and RTX 40 series, aiming to unlock the full potential of their Tensor cores. This modification makes consumer hardware significantly more capable for AI and compute-heavy workloads, narrowing the gap with professional-grade NVIDIA RTX workstation cards.
How the Mod Works
It is important to clarify that this mod is not intended for gamers. It will not increase frame rates in standard gaming scenarios, as the core graphics pipeline remains untouched. Instead, the focus is on AI-driven tasks, such as neural network inference and training. According to data from GitHub, the mod can deliver a transformative boost in Tensor core compute performance, ranging from 1.8x to 2x.
Technically, RTX Unlock does not require flashing a new VBIOS. Instead, it works by adjusting registers as the NVIDIA driver loads, bypassing the throughput restrictions imposed on consumer cards. This method also allows for the adjustment of power limits, voltage, and frequency offsets for both the GPU and video memory without the risks associated with traditional BIOS modification, though incorrect register settings could theoretically prevent the system from booting.
Tensor Core Performance Comparison
The actual performance gain varies depending on the architecture and the data format used. While most cards see a near-doubling of throughput, some models show different results:
- GeForce RTX 4090 (FP16/BF16): jumps from ~165 TFLOPS to ~330 TFLOPS.
- GeForce RTX 4070 Ti Super (FP16): increases from 95 TFLOPS to ~190 TFLOPS.
- GeForce RTX 2080 Ti (FP16): climbs from 46.9 TFLOPS to 93.7 TFLOPS.
However, the RTX 3090 Ti does not see a full doubling, showing gains of +80% for FP16 and +35% for BF16. Users should also keep in mind that real-world AI performance is often bottlenecked by memory bandwidth, not just raw core throughput.
Availability and Limitations
Currently, the mod is available only for Linux. This makes it a highly relevant tool for researchers and enthusiasts running local AI workloads on Linux-based systems, but it remains inaccessible for the majority of Windows users for the time being. While the mod is considered safe for the hardware, it operates at a low level, requiring caution when modifying system registers.
Гика
