For six years, the spec sheets have said it out loud and few people noticed. A GeForce card does tensor math with FP32 accumulation at half the rate it does the same math with FP16 accumulation. NVIDIA’s workstation cards built on the same silicon do not. This month a handful of modders found a way to switch the GeForce cards to the full rate, and the numbers they published confirm what the spec sheets implied: the hardware was always able to do it.
VideoCardz first reported the projects. We went to the repositories, the research paper behind them and NVIDIA’s own whitepapers, and checked the published numbers against NVIDIA’s specifications. The headline holds up for peak throughput. The real-world gains are smaller, and in some workloads they are zero.
NVIDIA’s own tables show the split
You don’t need a hack to see the cap. According to NVIDIA’s Ada Lovelace architecture whitepaper, the RTX 4090 does 330.3 dense FP16 tensor TFLOPS with FP16 accumulate, and 165.2 with FP32 accumulate. BF16 with FP32 accumulate is also 165.2. Ampere shows the same pattern in the GA102 whitepaper: the RTX 3090 Founders Edition is listed at 142 TFLOPS with FP16 accumulate and 71 with FP32 accumulate.
The same document lists the RTX A6000, a GA102 workstation card, at 154.8 TFLOPS for both FP16 and FP32 accumulate. No halving. It also lists the Turing-based TITAN RTX at 130.5 TFLOPS for both. So the half rate isn’t a law of physics or a property of the chip. NVIDIA gives the full rate to the cards it sells at workstation prices and halves it on GeForce.
FP32 accumulation is what you want when precision matters, which covers training and much of serious inference and image generation. FP16 accumulation is faster but loses precision, and many frameworks don’t use it by default. For AI work, the cap lands on the mode people actually use.
The unlock builds on a firmware bug found in a mining card
The root bug isn’t from a GeForce project at all. In June, Jon Pry published “A Canary in the Crypto Mine”. He describes a stack overflow in the signed booter firmware of NVIDIA’s CMP 170HX mining card that allows code execution in the GPU’s “Heavy Secure” mode, with “no signing key, no debug fuse, and no physical access.” His conclusion is the line that matters here: differentiation “the business model treats as hardware-permanent is, for the three axes we defeated, only firmware-gated.”
The GeForce projects build on the same line of work. According to the gpu-unlock repository by developer Zagitta, its Linux kernel module runs before NVIDIA’s driver loads, writes two “speed selector” registers and resets the card, so the unmodified driver then binds a card running at full rate. It doesn’t patch NVIDIA’s driver. The README itself credits earlier CMP unlocker work for the technique. A fork, gpu-unlock-ada, adds the RTX 4090 and says its AD102 payload and analysis come from another project, “TOTORONG/AD103_win_unlocker”, and that it “does not claim original discovery”. A third project, turing-unlock, has been tested on an RTX 2080 Ti.
Peak throughput doubles; real matrix multiplies gain less
On the RTX 4090, the claim of doubled peak throughput is backed by the authors’ own data. According to the gpu-unlock-ada test report, FP16 with FP32 accumulate went from 165.6 to 331.6 TFLOPS, and BF16 from 165.5 to 330.6. Those baselines sit almost exactly on NVIDIA’s 165.2 spec, and the unlocked numbers land on the 330.3 figure for FP16 accumulate. INT8, dp4a and plain FP32 didn’t move. That is consistent with a rate selector being flipped, not with a clock or power trick.
That is a synthetic instruction-rate test, though, and the report says it covers “one physical card and one successful cold power-on.” Real matrix multiplies tell a smaller story. According to the same repository, an independent owner’s matched run at a 450 W cap measured cuBLAS GEMM gains of 12 to 34% for FP16 and 16 to 20% for BF16 at two shapes, with BF16 “essentially unchanged” at 4096³. On a power-limited card, doubling the issue rate can mean hitting the power wall sooner.
Ampere looks better. According to the gpu-unlock project, an RTX 3090 Ti at 450 W went from 87.50 to 157.41 TFLOPS in a dense 8192³ BF16 GEMM, from 87.46 to 118.39 in FP16, and from 42.58 to 82.61 in TF32. On Turing, turing-unlock reports an RTX 2080 Ti at 250 W going from 52.8 to 77.9 TFLOPS of FP16 matrix multiply at 8192³, and from 55.7 to 91.4 at 2048³.
Local chatbots won’t generate text any faster
This is where the hype runs ahead of the evidence. Generating tokens in a local LLM is limited mostly by memory bandwidth, not tensor math, and the user reports show it. In the project’s own issue thread, one owner with four RTX 3090 cards capped at 240 W saw prefill go from 1,500 to 1,640 tokens per second, about 10%, with decode “unchanged within noise.” Another tested an RTX 3090 and an RTX 3060 with a Q8_0 quantized model in llama.cpp and measured no real improvement: prompt processing within 0.7% and token generation identical. Quantized models use integer paths that the unlock doesn’t touch.
Games won’t change either. Nothing in the projects’ data suggests a frame-rate gain, and none of the authors claims one. The benefit is real but narrow: FP16 and BF16 training, prompt processing on unquantized models, and diffusion work, on cards that have enough power headroom to use it.
The projects disagree on the risk
The projects don’t agree on how dangerous this is. The gpu-unlock project calls hardware harm “very unlikely” and says normal throttling and power management stay intact. The turing-unlock project warns that “tensor cores draw very high power” and that unlocking them “may damage your hardware.” This is unofficial code that executes an exploit in the GPU’s secure coprocessor at boot, on Linux, with a list of cards each validated on one or a few samples. Nothing stops NVIDIA from fixing the booter in a later driver, either. We found no statement from NVIDIA on the projects. If you depend on a card for work, this isn’t something to experiment with.
You bought the silicon, and NVIDIA rented you half of it
On this one point the evidence is strong. NVIDIA’s own whitepapers show full-rate FP32 accumulate on the RTX A6000 and TITAN RTX and half rate on the GeForce cards. Two RTX 4090 baselines read 165.6 and 160.4 TFLOPS, both close to NVIDIA’s half-rate 165.2 spec. Unlocked, they read 331.6 and 320.4, about twice as high and in line with the figure NVIDIA lists for FP16 accumulate. That likely isn’t a binning decision or a missing unit. It appears to be a setting, locked behind NVIDIA’s signed secure firmware. As we argued in How Nvidia Turned Gamers into Its R&D Budget, gamers paid for the Tensor Cores. They just didn’t get to use all of them.
Still, keep the claim in proportion. Double the peak isn’t double the speed. On a 4090 at stock power, the one independent matched run measured real gains of 12 to 34%, and there is nothing for most people running quantized chatbots. Segmentation is real. The bonus for most owners is modest. NVIDIA hasn’t said why GeForce gets half the rate, and as long as it doesn’t, owners will keep drawing their own conclusions.