Nvidia segments its GPU product lines not primarily by silicon capability but by firmware-enforced speed selectors that throttle instruction issue rates on consumer cards. The GA102, GA103, GA104, and GA106 chips inside GeForce RTX 3060 through 3090 Ti cards are physically capable of issuing FP16 and BF16 fused-multiply-add operations at double the rate Nvidia ships them at. A new open-source Linux kernel module called gpu-unlock now flips those selectors before the Nvidia driver loads, and the results are immediate: a 3090 Ti jumps from 87.5 to 157.4 TFLOPS in BF16, and TF32 nearly doubles from 42.6 to 82.6 TFLOPS. The module works by combining two previously known techniques — a speed-selector write from the CMP unlocker community and a pre-driver timing approach from UEFI firmware hacks — into a single kernel module that loads during initramfs, before the proprietary Nvidia driver binds the card. This eliminates the need to patch Nvidia's driver on every update, which was the main friction point with earlier approaches. The module writes SS0 (0x88888888) to select full BF16 rate and a 4-bit SS1 (0x8 IMLA4) selector, then performs a function-level reset. The entire unlock takes about one second per card. User-reported results across seven confirmed card models show BF16 dense GEMM gains ranging from +25% on power-limited 3090s (225W cap) to +80% on 3060 12GB cards. Real workload impact follows predictably: LLM prefill throughput gains of 14-36%, meaningful diffusion model speedups, and FlashAttention improvements. Memory-bandwidth-bound tasks like LLM decode and INT8 inference see no change, which is exactly what you'd expect when the bottleneck is instruction issue rate rather than memory bus. The engineering is notably careful. The module handles multi-GPU setups (up to four cards reported), display-driving GPUs, and VFIO passthrough to VMs. Suspend-to-RAM is handled with a notifier that re-applies the unlock before the Nvidia driver resumes. Users report bit-identical outputs before and after unlock — there is no precision compromise, only throughput. The project packages cleanly for Arch, Debian/Ubuntu, and Fedora/RHEL with proper initramfs integration across mkinitcpio, initramfs-tools, and dracut. The extraction story here is straightforward: Nvidia ships identical silicon to consumers and professionals, then enforces artificial throughput limits to protect margins on its Quadro and A-series professional cards. The price differential between a GeForce RTX 3090 and the professional equivalent it was segmented from is measured in thousands of dollars, not in transistors. This module reclaims capability the buyer already paid for in silicon. The obvious risk is Nvidia's response. The company could patch the speed-selector vulnerability in future firmware, could update driver initialization to detect and refuse unlocked cards, or could pursue legal action against the project (the generated firmware payload derives from pinned CMP reference firmware under Nvidia's terms). The project's GPL-2.0 licensing and transparent build process — downloading pinned Nvidia inputs over HTTPS with SHA-256 verification — are defensible engineering choices but not legal shields. For the broader AI/ML ecosystem, this matters because RTX 30-series cards remain the most common consumer GPUs used for local inference and fine-tuning. Doubling BF16 throughput on hardware that millions of people already own is a pure efficiency gain with zero additional resource consumption — same wattage, same cooling, same silicon. It is the cleanest possible example of generativity: unlocking value that was artificially locked.
