A Linux Module Lifts the BF16 Cap on GeForce RTX 3000 Cards. Power Limits Decide the Gain

A kernel module published on GitHub by a developer who goes by Zagitta says it restores the full instruction issue rate for FP16 and BF16 fused multiply-add operations on GeForce RTX 3000 cards, and its own benchmark shows an RTX 3090 Ti rising from 87.5 to 157.4 TFLOPS in BF16 matrix multiplication at 450 W. The project, gpu-unlock, reached 153 points and 59 comments on Hacker News on October 24, 2026.

The number that matters here is not 157.4 but 450. Every gain the project reports is a function of the power limit the card is allowed to draw, and the cards most people own do not run at 450 W.

What the module does

The repository describes gpu-unlock version 1.0.0 as a Linux kernel module under GPL-2.0-only that targets the GA102, GA103, GA104 and GA106 chips. It runs before the nvidia driver binds the card, so no driver patching is needed. According to the README, it writes two speed selectors (SS0 0x88888888, SS1 0x8) and then performs a function-level reset.

The project's claim is that it "doubles the instruction issuing speed of FP/BF16 FMA into FP32 operations" and brings the cards level with their professional equivalents. That framing is the author's. Hacker News commenter RachelF put the same point more bluntly: "The hardware supports it, but then Nvidia disabled it if the driver detects cheaper cards." Neither statement comes from Nvidia, and the README contains no warranty language.

The method was clarified in the thread after commenter superkuh disputed the headline's "16 bit" framing. Zagitta replied that the change concerns FP/BF16 FMA instructions accumulating into FP32, which is the path matrix engines use. Zagitta also wrote that the technique is a DMA buffer overflow with return-oriented programming, so that "no unsigned code is ever executed". Commenter londons_explore expects future firmware to close it.

The one controlled benchmark

The only controlled measurement in the README is the author's own: a dense 8192³ GEMM on an RTX 3090 Ti at 450 W, FP32 accumulation, median of five samples. It is a single-author, single-machine result. Other benchmark claims on this site are covered in the Tov comparison and the Roc deflate piece.

Input type	Before (TFLOPS)	After (TFLOPS)
FP16	87.46	118.39
BF16	87.50	157.41
TF32	42.58	82.61

The README says the per-card gains below come from each reporter's own benchmark and "are not directly comparable". It adds that the lowest figure, +25%, came from cards capped at 225 W.

1. Before87.5 TFLOPS
2. After157.41 TFLOPS

Source: gpu-unlock README (author-run), accessed 2026-10-24

Reported gains by card

The README's compatibility table lists the following BF16 GEMM gains. Report counts are small: the RTX 3090 has seven, the 3080 20 GB has two, and the 3060 12 GB has two.

Card	Chip	BF16 GEMM gain
RTX 3090 Ti	GA102	+68 to 80%
RTX 3090	GA102	+25 to 67%
RTX 3080 10 GB	GA102	+70 to 73%
RTX 3070	GA104	+73 to 81%
RTX 3060 12 GB	GA106	about +80%
RTX 3080 Ti 20 GB (OEM)	GA102	none; BF16 stays at half rate

The OEM 3080 Ti 20 GB completes the procedure but gains nothing. The RTX 3070 Ti was not tested with the module, though the README says the same change worked on it through a UEFI port on Windows. GA103 has no reports.

What users measured

The Hacker News thread is the nearest thing to independent testing, and it is uneven. Commenter Tiberium wrote "I've tested myself and the results are legit" and reported up to 40 to 50% in some workloads at the same power limit. Commenter aunty_helen reported about 133 tokens per second on a 27B model, "about 3.2x". Commenter jacquesm reported a 30% cut in wall-clock time on one setup, but said that with seven RTX 3090s they "all require a power increase to get > 25% gain".

At lower caps the picture changes. Commenter chromaton saw about 6% faster training on an RTX 3060 limited to 170 W. Commenter blagui wrote that "the real gains are far less than the advertised", citing about 25% at a 200 W cap and roughly 7 to 8% in practice. These are forum posts, not published benchmarks, and none states a batch size or framework version.

The workload matters as much as the wattage. The README says matrix-heavy jobs speed up: LLM prefill by 14 to 36% in some reports, plus diffusion and FlashAttention. LLM decode is limited by memory bandwidth and stays about the same, as does INT8. Reports that compared outputs found them bit-identical.

What would change the read

The cost side is thermal. Jacquesm wrote that users "will need to run your card hotter than you might be comfortable with" and that the gains need up to 450 W, "well above what the stock cooler was designed for". Zagitta said the module has run for a month on two water-cooled RTX 3090s with no issues. Commenter chromaton said "I have no desire to brick my card or set it on fire".

The README lists operational risks too: a firmware hang requires a full power-off before retrying, the display card goes blank during the reset, and suspend-to-RAM resets the selectors, which the module re-applies at a cost of about one second. Hibernation is not covered. Setup means editing the kernel command line with gpu_unlock.targets=all and rebuilding the initramfs.

Scope is narrow. Zagitta wrote that Hopper and Blackwell have "a very different security chain" and are out of scope for now. A separate Hacker News submission the same day, a source release for the RTX 4090, had 3 points. For anyone running a 3090 at stock power on an inference box, decode throughput will not move. A user who raises the limit toward 450 W and runs prefill-heavy jobs is the case the 157.4 figure describes. What would falsify the strong reading is a repeat of the 8192³ test on a 3090 at 350 W with a named framework and batch size; the README does not contain one.

The next fixed point is the compatibility form the README offers, where reports for the 3070 Ti and GA103 are still missing.