RTX 20, 30 and 40 GPUs Unlock Higher AI Matrix Throughput

| 2 Min Read
Unofficial firmware tools lift Nvidia’s half-rate AI math limit on select GeForce cards, with matrix gains varying by model, power cap and workload.

Unofficial firmware tools are exposing more AI matrix throughput on select Nvidia GeForce RTX 20, 30 and 40-series cards. The gains matter most for workloads that use matrix math, such as prompt processing and image generation; they don’t translate into a blanket doubling of application speed. In particular, chatbot token generation remains roughly unchanged because it depends heavily on memory bandwidth.

The reported unlock lifts a rate limit that Nvidia’s published specifications apply to affected GeForce cards: FP32-accumulate performance runs at half the rate of FP16-accumulate. Turing and Ampere professional cards aren’t subject to the same limit. Removing it can raise peak tensor rates sharply, but benchmark results show that real matrix multiplication improves by a more varied amount.

Support depends on the GPU and the project

Three Linux projects target different GeForce generations, while experimental Windows builds use another route. Their compatibility lists are uneven: a reported result on one card doesn’t establish that every board in the same generation will work.

Which GeForce cards each unlock covers

Generation

Project

Cards with reported results

Experimental or untested

RTX 20 (Turing)

turing-unlock

RTX 2080 Ti (22GB memory-modded cards)

Other Turing GPUs, added untested; FP16 only, since Turing lacks BF16 support

RTX 30 (Ampere)

gpu-unlock

RTX 3090 Ti, 3090, 3080 10GB, 3080 20GB (memory-modded), 3070, 3060 12GB

RTX 3070 Ti and GA103 cards untested; one RTX 3080 Ti 20GB OEM card stays at half rate

RTX 40 (Ada Lovelace)

gpu-unlock-ada

RTX 4090

RTX 4090 D, 4070 Ti Super, 4060 Ti 16GB (experimental, one card each)

RTX 50 (Blackwell)

None

—

Not supported; gpu-unlock’s developer calls it out of scope

For Ampere, gpu-unlock lists GA102, GA103, GA104 and GA106 series IDs. Its reported results include the RTX 3090 Ti, 3090, 3080 10GB, 3070 and 3060 12GB, along with a memory-modded 3080 20GB. A 3070 Ti and GA103 cards remain untested; one RTX 3080 Ti 20GB OEM model reportedly stays at half rate.

The Ada-focused gpu-unlock-ada fork reports results on the RTX 4090. It lists the 4090 D, 4070 Ti Super and 4060 Ti 16GB as experimental, with one card tested for each. Turing-unlock, adapted from gpu-unlock, has a reported result for a memory-modded RTX 2080 Ti with 22GB. Other Turing models appear in its compatibility list, but haven’t been tested. Turing also lacks BF16 hardware, so its reported gains cover FP16.

Peak throughput and workload gains tell different stories

The RTX 4090 figures illustrate the gap between peak capacity and application performance. According to the Ada fork’s developer, unlocking raises its FP16 tensor rate from 165.6 to 331.6 TFLOPS and its BF16 rate from 165.5 to 330.6 TFLOPS. That’s close to a doubling in peak throughput. In one owner’s cuBLAS GEMM tests, however, median gains were 12–34% for FP16 and up to 20% for BF16.

A separate Windows test on an MSI RTX 4070 Ti found larger matrix-multiply gains: 80.1% for FP16 and 76.4% for BF16. Power limits may help explain why results differ between tests. The project’s data isn’t a controlled comparison across cards, and its README cautions that user submissions aren’t directly comparable.

A modified-driver test on the RTX 3090 Ti reported improvements of 35% for FP16, 80% for BF16 and 94% for TF32 matrix multiplication. Other Ampere user reports put BF16 gains at 25–67% on the RTX 3090, 70–73% on the 3080 10GB, 56–59% on a memory-modded 3080 20GB, 73–81% on the 3070 and about 80% on the 3060 12GB. The lowest RTX 3090 result came from a card capped at 225 watts, underscoring how power limits can shape the outcome.

For AI workloads, Ampere LLM prefill improved by 14–36%, and diffusion and FlashAttention also ran faster in reported tests. LLM decode and INT8 performance stayed “about the same.” That split is useful context: a faster matrix path can shorten some stages of inference without speeding up every part of the pipeline.

Turing results come from modified RTX 2080 Ti cards with 22GB of memory. A Dell OEM model limited to 250 watts showed FP16 matrix-multiply gains from 36% to 64%, depending on matrix size; another card running at 260 watts gained 59%. KohakUnlock testing on a Leadtek 2080 Ti 22GB reported an 85% improvement on an instruction-rate test and a best matrix-multiply median gain of 53% against a locked card of the same model. One separate pre-boot loader test on a stock 11GB ASUS RTX 2080 Ti reported a 77% FP16 gain.

Matrix-multiply benchmark gains, unlocked vs. stock

Card

Precision

Gain

Tested by

RTX 4090

FP16 / BF16

+12–34% / up to +20%

One owner, Linux, 450 W

RTX 4070 Ti (MSI)

FP16 / BF16

+80.1% / +76.4%

Separate Windows port, one card

RTX 3090 Ti

FP16 / BF16 / TF32

+35% / +80% / +94%

One user, modified Nvidia driver, 450 W

RTX 3090

BF16

+25–67%

User reports; lowest at a 225 W cap

RTX 3080 10GB

BF16

+70–73%

User report

RTX 3080 20GB mod

BF16

+56–59%

User reports

RTX 3070

BF16

+73–81%

User reports

RTX 3060 12GB

BF16

About +80%

User reports

RTX 2080 Ti 22GB mod (Dell)

FP16

+36–64%

turing-unlock developer, 250 W, by matrix size

RTX 2080 Ti 22GB mod

FP16

+59%

One user, 260 W

RTX 2080 Ti 22GB mod (Leadtek)

FP16

+53%

KohakUnlock; best median vs. a locked card

RTX 2080 Ti 11GB (ASUS, stock)

FP16

+77%

haohervchb pre-boot loader

These figures describe specific matrix and instruction tests, not a general 2× boost to applications or games. The practical case is narrower: workloads that can use the unlocked math rate may finish certain compute-heavy steps sooner, while memory-bound stages can see little change.

The firmware path and its limits

The gpu-unlock explainer traces the cap to two “speed selector” registers that the host driver can’t write directly. Nvidia’s signed SEC2 “Booter” firmware, however, contains a stack buffer overflow that can be used to run code in “heavy secure” mode. The exploit became possible after Nvidia published a debug Booter encrypted with a publicly known test key.

The module uses a four-byte write with the V67 payload to open the relevant privilege mask, allowing the host to select the full rate. It then resets the GPU before Nvidia’s driver attaches. The underlying stack overflow was documented by Jon Pry in “A Canary in the Crypto Mine: Defeating Stack Protection in a GPU Secure Coprocessor,” published on Zenodo in June. Pry found the issue in the Heavy Secure Falcon on the CMP 170HX. The V67 constants came from WildFlash1st’s cmp90hx-unlock project, while gpu-unlock’s early boot timing appears to draw on that developer’s separate Windows EFI bootloader.

The CMP Unlocker, which restores locked VRAM on the CMP 170HX, is another known use of the exploit. Zagitta, gpu-unlock’s maintainer, said on Hacker News that the project is their work, while “the exploits … aren’t my ideas by any means.” They described the GeForce change as a matter of flipping “the 2 bits” needed to select the higher rate.

Compatibility remains a moving target. The Ada fork says unverified boards shouldn’t be treated as supported just because they share firmware with a tested model; turing-unlock also identifies untested device IDs. The projects are collecting owner reports, but gaps remain across Turing and Ada, as well as GA103 Ampere cards. RTX 50-series Blackwell GPUs aren’t supported, and gpu-unlock’s developer describes that generation as outside the project’s scope. The useful takeaway for AI builders is to judge the unlock by their exact card and workload: peak tensor figures set an upper bound, while board limits and memory traffic decide how much speed reaches real work.

The project READMEs also warn that firmware modifications can hang a GPU or host, prevent boot, corrupt data or damage hardware. Those risks sit alongside the uneven card coverage, making the published measurements most useful as evidence of potential rather than a promise of repeatable results.

Related reporting: VideoCardz reported the discovery. Background on the tested cards and related hardware includes the RTX 4090, the RTX 3090 Ti, Turing and Ampere pro cards, LLM prefill, the 22GB RTX 2080 Ti memory mod and the CMP Unlocker.

Source: Shane Downing · www.tomshardware.com

Comments

Please sign in to comment.