Modders Unlock AI Compute Limits on NVIDIA RTX 20, 30, and 40 Series GPUs

NVIDIA GeForce RTX 20, 30, and 40 series graphics cards can now achieve up to a twofold increase in Tensor Core throughput through community-developed firmware and low-level configuration tools, increasing performance for local artificial intelligence workflows.

The Tech TL;DR:

• Modders removed firmware-level Tensor Core compute restrictions on consumer NVIDIA RTX 20, 30, and 40 series GPUs, achieving nearly double the throughput for specific matrix math operations.
• Real-world performance gains vary by software and workload; while FP16 and BF16 tensor throughput spikes significantly, local LLM token generation speeds remain bottlenecked by memory bandwidth.
• The exploits operate by changing low-level configuration registers or exploiting bootloader vulnerabilities before NVIDIA drivers load, though the modifications void hardware warranties and lack support for the RTX 50 Blackwell series.

Tools Remove Tensor Core Rate Limits on NVIDIA GPUs

The research builds on earlier attempts to unlock restricted compute capabilities on NVIDIA CMP mining GPUs. Developer Zagitta adapted those techniques for GeForce RTX 30 cards through an open-source project, while work involving firmware extended support to the Ada architecture. Instead of modifying graphics drivers, the tools change low-level GPU configuration registers before the NVIDIA driver loads, removing limits on the rate at which certain Tensor Core instructions can be issued. Tom’s Hardware noted that the explainer for gpu-unlock states that the rate limits reside in two speed selector registers that the host driver cannot write, but NVIDIA’s signed SEC2 Booter firmware contains a stack buffer overflow exploited to run code in heavy secure mode.

Both unlocks successfully pass NVIDIA’s driver checks and override PCIe Function Level Reset, running stock NVIDIA drivers while CUDA benchmarks execute normally. According to TweakTown, an RTX 3090 Ti at its 450W power limit went from 87.46 to 118.39 TFLOPS in FP16, from 87.50 to 157.41 TFLOPS in BF16, and from 42.58 to 82.61 TFLOPS in TF32. On the RTX 40 series, the RTX 4090 moved from 165.6 to 331.6 TFLOPS on FP16 and 165.5 to 330.6 TFLOPS on BF16, reaching performance figures typically reserved for professional cards like the NVIDIA RTX 6000 Ada.

Expanding Turing Architecture Support

The modification scope expanded to include the older GeForce RTX 20 series. TechPowerUp updated its reporting to confirm that NVIDIA’s RTX 20 series Turing cards are unlocked using the same technique, doubling performance metrics. TweakTown documented that an RTX 2080 Ti reached 111.0 TFLOPS in an FP16-to-FP32 test, compared with 59.9 TFLOPS prior to the compute unlock. That particular card was physically modified with 22GB of VRAM, which is distinct from the software unlock mechanism.

Independent test results compiled across multiple repositories indicate that real-world application speedups depend heavily on the targeted framework. HotHardware highlighted that an RTX 3090 achieved 73.0 to 122.8 TFLOPS in a PyTorch GEMM test. However, unsupported configurations frequently fail to boot, and all modifications require manual application at every system boot because the changes clear when the GPU loses power.

Application Gains Vary by Workload and Memory Bandwidth

While peak synthetic tensor rates double across several architectures, application-level gains vary significantly. TweakTown reported that an RTX 3090 Ti running ComfyUI observed a 56% to 59% increase in throughput for BF16 image models, whereas INT8 workloads remained largely stationary. For local large language models, prompt processing improved by 14% to 36%, but token generation speeds remained unchanged because generation is bounded by memory bandwidth rather than raw compute capacity.

Hardware analysts note that while these firmware exploits narrow the performance gap between consumer desktop cards and enterprise accelerators, they do not replicate the massive VRAM capacity or full software stack support of professional hardware. Modifying GPU firmware voids manufacturer warranties. As developer documentation indicates, the newer GeForce RTX 50 Blackwell series remains entirely unsupported and out of scope for current open-source unlock utilities.

More on NVIDIA RTX: NVIDIA RTX Spark Laptops Arrive Worldwide with Blackwell Graphics and AI Platform