The Locked Silicon

A GPU owner buys the die, the memory, and the power delivery. The vendor keeps one thing: a signed file that decides how fast the tensor cores are allowed to count. This report documents the best-measured case of feature locking in consumer hardware—and why it matters to anyone running local inference.

Same Chip, Two Price Tags

Since the introduction of tensor cores, NVIDIA has segmented its lineup not only by memory size and power budget, but by an instruction issue rate enforced in software. On consumer GeForce RTX 20, 30, and 40 series cards, FP16 and BF16 tensor operations that accumulate in FP32—the accumulation mode used by most training and a meaningful share of inference work—execute at roughly half the rate of the identical operation on the company's professional cards built from the same silicon families.

The restriction is not a physical property of the die. Professional parts of the same architectures (Quadro RTX, RTX A-series, A40, A100) perform the identical instruction at full rate. The difference ships in the driver, and it follows the consumer card wherever it goes.

The Measurement

In late 2026, a community-written modification targeting the driver's instruction issue limitation was published with reproducible benchmarks. A single kernel module adjusts the issue rate for FP16/BF16 instructions with FP32 accumulation before the driver binds—no vBIOS flash, no driver patching, and no re-application needed after driver updates. Published dense matrix multiplication (GEMM) results for an RTX 3090 Ti operating at a 450 W power limit:

The hardware was never slower. It was permitted to be slower, at a rate calibrated to protect the price ladder above it. Coverage of the unlock appeared in October–November 2026 from multiple independent hardware outlets, with the measurement numbers originating from the project's own published benchmark suite.

This Is Not New Practice

Instruction-rate locking is the newest chapter of a long pattern. In December 2017—the same month the Titan V launched—NVIDIA amended its GeForce and Titan driver license terms to prohibit datacenter deployment of those products, a restriction enforced not in silicon but in a click-through agreement. The Quadro-versus-GeForce division has operated for two decades on the same principle: similar dies, differentiated by firmware, certification, and driver features such as ECC memory and certified OpenGL paths.

Note what all of these mechanisms have in common: the vendor limits its own product's capabilities through channels it controls—its signed driver, its license text, its feature flags. Whether one calls this segmentation, product management, or planned constraint is a matter of economics. That it happens is not in dispute; it is measurable with a stopwatch and a matrix multiply.

Why It Matters for Local Inference

For gaming workloads the lock is largely invisible. For AI workloads it is substantial: prompt processing, fine-tuning, and diffusion models lean heavily on exactly the penalized precision/accumulation combinations. A developer who bought a top-tier consumer card for local inference has been running at a fraction of the silicon's measured capability, on a schedule set by the vendor's pricing strategy rather than the physics of the chip.

The older lesson holds: professional-class cards of previous generations—the workhorse hardware of the current local-AI wave—were never subject to these particular locks. Much of the fleet running local inference today owes its throughput to having been sold to a different market segment before the restrictions existed.

What Owners Should Know

• The lock affects FP16/BF16 with FP32 accumulation, and some TF32/FP8 paths, on consumer RTX 20/30/40 series. FP4 workloads are unaffected on current parts.
• Gaming performance is essentially unchanged; training and some diffusion/prompt-processing workloads are the affected classes.
• Community unlock tooling exists; it modifies how the driver binds, carries the usual risks of unofficial modification, and may affect stability, thermals, and warranty standing. Measure your own workloads before and after; do not take any benchmark, including ours, on faith.
• Professional-class and datacenter cards of the same architectures are not subject to these particular rate locks.