Success reports: post working unlocks here

[Zagitta]

Did gpu-unlock work on your card? Post it here as a comment instead of opening an issue.

Please include:

• Card: brand, model and memory size, e.g. EVGA RTX 3090 FTW3 24 GB
• PCI ID: from lspci -nn -d 10de:, e.g. 10de:2204
• Laptop or desktop
• Distribution, kernel and NVIDIA driver version
• Before/after numbers if you have them: tool, matrix size, accumulation type, power limit, and TFLOPS for FP16, BF16 and TF32. Real workloads (LLM prefill, training, image generation) are welcome too.

Keep it short. Logs aren't needed when it worked.

If the unlock failed, gave no speedup, or stopped working later (after idle, suspend or a driver reload), open an unlock failure issue instead.

Earlier success reports were filed as issues and have been closed in favour of this thread. They are still searchable under closed issues.

[aucomp]

gpu-unlock 1.1.0

Card: MSI RTX 3090 Suprim X (1462:3882, VBIOS 94.02.42.00.F5), power limit set to 350 W.
Model: qwen3.8:27b-cline (Ollama, Q4_K_M, q8_0 KV, 350 W cap)
NVIDIA Driver 580.178.04
Ubuntu 24.04, kernel 6.8.0-147
PCI ID: 10de:2204
Dedicated desktop

Test	Before	After	Change
Writing, edits (json)	94.6 tok/s	94.6 tok/s	none
Writing, prose	49.6 tok/s	49.4 tok/s	none
Reading 15.8k tokens	1201 tok/s	1219 tok/s	+1.5%
Reading 48k tokens	45.2 s	43.6 s	+3.7%
Reading 96k tokens	109.1 s	102.4 s	+6.6%
Reading 111k tokens	134.5 s	125.3 s	+7.4%
Max GPU temp	67 °C	67 °C	none

Writing speed is limited by memory bandwidth, so only prefill (FP16 flash attention) gains, and the gain grows with context.

An unquantized model may have more noticeable speed improvements.

[Path-Seeker]

Result: Unlocked, and FP16/BF16 got faster

Card: ASUS ROG Strix GeForce RTX 4080 16 GB, non-OC (subsystem 1043:88c0)

PCI IDs (lspci -nn -D -d 10de:)

0000:01:00.0 VGA compatible controller [0300]: NVIDIA Corporation AD103 [GeForce RTX 4080] [10de:2704] (rev a1)
0000:01:00.1 Audio device [0403]: NVIDIA Corporation AD103 High Definition Audio Controller [10de:22bb] (rev a1)

VBIOS version: 00000000:01:00.0, 95.03.2B.00.CE

Is the card modified? Stock

Distribution: Kubuntu 26.04

Kernel: 7.0.0-38-generic

NVIDIA driver: 580.178.04, proprietary kernel modules (Ubuntu's prebuilt linux-modules-nvidia-580)

Initramfs generator: initramfs-tools

gpu-unlock version and install method: PR #24 at 9bf4cb5 (1.2.0), make install (install-deb)

Kernel command line

BOOT_IMAGE=/boot/vmlinuz-7.0.0-38-generic root=UUID=e6307b4a-0763-487d-a1b2-9b224d1ed51b ro kvm.enable_virt_at_load=0 gpu_unlock.targets=0000:01:00.0

Setup: single GPU, which is also the display card (X11, SDDM, KDE Plasma); iGPU disabled. Stock 320 W power limit. Secure Boot off. ASUS ROG STRIX Z690-F GAMING WIFI, i7-13700K. The unlocked boot was a one-off edit of the kernel line in GRUB.

Kernel log (journalctl -k -b | grep -i gpu_unlock, plus the selectors line)

gpu_unlock: loading out-of-tree module taints kernel.
gpu_unlock: module verification failed: signature and/or required key missing - tainting kernel
pci 0000:01:00.0: gpu_unlock: stage=claim device=0x2704 family=AD102 cmd=0x3 original=0x3 bar_len=0x1000000 region=1 enabled=0 master=0
pci 0000:01:00.0: gpu_unlock: stage=layout fb=0x3ff800000 vga=0x3ff700000 frts=0x100000 wpr=0x3f4c00000..0x3ff700000 boot=0x3ff5f7000 gsp=0x3fa5a0000 heap=0x3f4d00000
pci 0000:01:00.0: gpu_unlock: stage=dma booter=56832 payload=0xfa00 pages=20606 meta=0xfffdf000 sig=0xfffe0000 stock=0xfffde000 boot=0xfffc0000
pci 0000:01:00.0: gpu_unlock: stage=gfw plm=0xffffff8f ss0=0x65543074 ss1=0x3 wpr2_hi=0x0
pci 0000:01:00.0: gpu_unlock: stage=fwsec version=3 fuse=0x1 ver=1 scratch=0x0 wpr2_raw=0x3ff6000:0x3ff6e00 frts=0x3ff6000:0x3ff6e00
pci 0000:01:00.0: gpu_unlock: stage=publish meta=0xfffdf000 sig=0xfffe0000 size=0xfa00 radix=0xfffbf000
pci 0000:01:00.0: gpu_unlock: stage=host ss0=0x88888888 ss1=0x8 plm_feat=0xffffffff plm_wpr_cfg=0xfffff0ff
pci 0000:01:00.0: gpu_unlock: stage=flr post_wpr2=0x1ffffe00:0x0 ss0=0x88888888 ss1=0x8 plm=0xffffffff cmd=0x3 devsta=0x0 chip=0x193000a1
pci 0000:01:00.0: gpu_unlock: stage=handoff ret=0 state=4 retained=0 pages=0/20606 cleanup=0
pci 0000:01:00.0: gpu_unlock: unlocked
pci 0000:01:00.0: AD10x selectors-applied: readout1=0x0 full=1

nvidia bound the card normally afterwards: desktop up, no Xid or NVRM errors.

Benchmark before and after

torch 2.10.0+cu128 (cuBLAS via torch.matmul), dense 8192³ GEMM, FP32 accumulation (TF32 with allow_tf32=True). Each figure is the median of 20 runs after 3 warm-ups, at the stock 320 W limit. The before and after runs were separate boots, without and with gpu_unlock.targets=.

Input	Before	After	Gain
FP16	104.1 TFLOPS	168.2 TFLOPS	+62%
BF16	104.1 TFLOPS	177.0 TFLOPS	+70%
TF32	53.0 TFLOPS	83.5 TFLOPS	+58%

Numerical parity: the SHA-256 of a seeded 4096³ product is identical before and after for FP16, BF16 and TF32.

Real workload: BF16 training of a 196M-parameter Transformer (8 encoder / 2 decoder layers, d_model 1024, ~80k tokens per optimizer step, quickmt-train).

• Stock: ~1.22 s/step, ~58k tok/s.
• Unlocked: ~0.83 s/step, ~85k tok/s at steady state (30 min run), about 1.45× end to end.
• Loss continued smoothly from the stock checkpoint.

Suspend and resume: Not tested

Anything else: an independent data point for 10de:2704 next to the +53% one in the README table. Here the card held 320 W without raising the power limit, and BF16 came out at +70%.

[chadcox]

gpu-unlock 2.0.0: RTX 4070 (AD104, 10de:2786) works

• Card: ASUS RTX 4070 12 GB (subsystem 1043:88e1), VBIOS 95.04.3E.80.73, 200 W stock power limit
• PCI ID: 10de:2786
• Desktop
• OS: Fedora 44, kernel 7.2.7-200.fc44, NVIDIA 615.71.09, installed with make install-rpm (DKMS)

PyTorch 2.11 (cu128), 8192³ GEMM, FP32 accumulation (reduced-precision reduction off), 200 W limit:

Test	Before	After	Change
FP16 GEMM	61.7	91.7	+49%
BF16 GEMM	62.1	93.5	+51%
TF32 GEMM	30.1	43.7	+45%
FP32 GEMM (control)	19.0	19.1	—
BF16 SDPA causal	57.2	85.2	+49%

TFLOPS. The card is power-bound: it holds about 199 W in both runs, and the clock drops from 2655 to 2325 MHz after the unlock. A higher power limit would probably bring it closer to the 4070 Ti / 4070 SUPER results. No Xid errors.

[jbrewer99x]

gpu-unlock 2.0.0: RTX 3060 12 GB LHR (GA106, 10de:2504) on TrueNAS SCALE

• Card: ASUS RTX 3060 12 GB LHR (subsystem 1043:881d), VBIOS 94.06.2F.40.59, 170 W stock power limit
• PCI ID: 10de:2504
• Desktop (home server)
• OS: TrueNAS SCALE 25.10.7, kernel 6.12.105-production+truenas, NVIDIA 580.173.02
• Install: the Debian packages don't fit TrueNAS (read-only root, nvidia loaded by middlewared from a sysext), so I built the module off the box and loaded it from a middlewared.service drop-in, see Add TrueNAS SCALE support #240.

LocalAI (ggml backends), medians of 3 runs (2 for music), two stock runs to show noise:

Test	Before	After	Change
Trellis 2 4B, 512 game mesh	142.9 / 142.3 s	138.1 s	+3.0–3.5%
Z-Image-Turbo 1024², 25 steps	62.5 / 62.5 s	61.2 s	+2.0%
FLUX.2 klein 4B 1024², 4 steps	7.50 / 7.54 s	7.30 s	+2.7–3.3%
ACE-Step turbo, 30 s music	6.39 / 6.32 s	6.03 s	+4.8–5.8%
Stable Audio, 5 s	0.264 / 0.267 s	0.257 s	+2.7–3.7%
Qwen3 TTS	4.97× / 4.96×	5.05×	+1.6%
LLM generation, 4B and 9B Q4	~70 tok/s	~71 tok/s	within noise
LLM prefill ~3k, Q4	3364 / 2946 tok/s	3406 tok/s	inconclusive

Small but consistent gains on diffusion/flow/audio; no synthetic GEMM numbers.

[vcgh1]

gpu-unlock 1.0.0 (990a14e): RTX 3070 (GA104, 10de:2484) works, Secure Boot on

• Card: ZOTAC RTX 3070 Twin Edge OC 8 GB, VBIOS 94.04.3A.40.67, 220 W stock power limit
• PCI ID: 10de:2484
• Desktop: Gigabyte B550I AORUS PRO AX, Ryzen 7 5800X3D
• OS: Ubuntu 26.04.1, kernel 7.0.0-38-generic, nvidia-driver-595-open 595.99.02; DKMS debs
• Secure Boot on: Ubuntu's DKMS signed the module with the enrolled MOK; only taint 4096 (out-of-tree)

PyTorch 2.14.1, 8192³ GEMM, FP32 accumulation (torch defaults), 220 W, stock clocks:

	Before	After	Change
FP16 TFLOPS	45.03	76.96	+70.9%
BF16 TFLOPS	44.55	75.01	+68.4%
TF32 TFLOPS	22.33	37.73	+69.0%
INT8 TOPS (control)	35.24	35.24	0%

The card stays on its 220 W cap (GEMM clock 1920 → 1650 MHz), hence ~1.7×.

Real workloads: vLLM 0.31.0 Qwen2.5-1.5B BF16 prefill +58–59%; ComfyUI SDXL 1024², 30 steps, 13.36 → 9.23 s/image with bit-identical images; Ollama Qwen2.5-7B Q4_K_M prefill +3–8%, decode unchanged. Seeded GEMM hashes and temperature-0 LLM outputs were bit-identical, and a 2 h soak (235k GEMMs) found no mismatch.

Also unlocks after a warm restart from Windows (Windows NVIDIA driver loaded) straight into Linux on this board.

[wetakush]

gpu-unlock 2.0.0 (1fc6d62) + one-line patch: RTX 3050 8 GB desktop (GA107, 10de:2582) works

• Card: RTX 3050 8 GB, GA107 revision (subsystem 10de:2582), VBIOS 94.07.68.00.5C, 115 W stock power limit, max SM clock 2100 MHz
• PCI ID: 10de:2582
• Desktop, the card drives the display (Hyprland, Wayland)
• OS: Arch Linux, kernel 7.2.9-arch1-1, nvidia-open-dkms 615.78.08; packages from make arch, mkinitcpio UKI + systemd-boot, Secure Boot off

GA107 isn't in recipes.c, so I added it to the ga10x family, the same way the GA107 laptop reports did (#95, #113):

#define GA107(id) { id, 0x177, &ga10x }
...
	GA107(0x2582),

No other changes. PMC_BOOT_0 reads 0xb77000a1 (chip 0x177), SS0/SS1 go from 0x41004156/0x3 to 0x88888888/0x8, and the log ends with unlocked.

PyTorch 2.14.1 (cu130), 8192³ GEMM, FP32 accumulation (reduced-precision reduction off), 115 W, median of 5:

	After
FP16 TFLOPS	29.32
BF16 TFLOPS	28.75

No stock run yet. The half-rate ceiling for this card is 20 SMs × 512 FLOP/clk × 2.1 GHz = 21.5 TFLOPS, so 29.3 is above anything it can reach while locked. Ollama was holding 6.3 GB of VRAM during the run, so TF32 at 8192³ ran out of memory and is missing.

[PeishuenWu]

Card: NVIDIA GeForce RTX 3060 Mobile / Max-Q 6 GB (10de:2520, Subsystem 1043:1722 ASUS ROG Zephyrus G14 GA401QM, 60W)
OS / Kernel / Driver: Ubuntu 24.04.5 LTS (6.8.0-124-generic), NVIDIA Driver 595.99.02
Fix: Added GA106(0x2520) to kernel/gpu_unlock/recipes.c (confirming #212)

**8192x8192 GEMM Benchmark (FP32 accumulation, 60W stock power limit):**
- **FP32:** 5.39 TFLOPS -> 5.39 TFLOPS (0%)
- **TF32:** 10.12 TFLOPS -> 14.94 TFLOPS (**+47.6%**)
- **BF16:** 20.06 TFLOPS -> 29.05 TFLOPS (**+44.8%**)
- **FP16:** 19.31 TFLOPS -> 26.72 TFLOPS (**+38.4%**)

Kernel log:
`pci 0000:01:00.0: gpu_unlock: stage=flr post_wpr2=0x1ffffe00:0x0 ss0=0x88888888 ss1=0x8 plm=0xffffffff cmd=0x2 devsta=0x0 chip=0xb76000a1`
`pci 0000:01:00.0: gpu_unlock: unlocked`

[gamr416]

RTX 3060 12GB — Forge SD + Bonsai-2 GGUF

Hardware / software

• GPU: NVIDIA GeForce RTX 3060 12GB (GA106, PCI 0000:01:00.0)
• Driver: 615.78.08
• OS: Arch Linux, kernel 7.2.11-arch1-1
• Boot: systemd-boot + gpu_unlock.targets=all
• gpu-unlock: 2.0.0 (DKMS + firmware via make install-arch)
• Unlock status: dmesg shows gpu_unlock: unlocked, host SS0=0x88888888 / SS1=0x8

Image gen (Stable Diffusion Forge)

• Model: Illustrious-based checkpoint, ~6GB .safetensors
• Settings: 1024×1024, 25 steps, Euler a, CFG 7, batch 1
  (1 warmup + 3 timed runs, wall-clock via API)

	Before	After
Avg time / image	22.8 s	14.4 s
Approx. it/s	1.09	1.74

~1.59× throughput (−37% time per image)

LLM (Ternary-Bonsai-2 27B PQ2_0 GGUF, PrismML llama-bench)

• Prefill 2000 tokens / decode 128 tokens
• -ngl 99, KV cache q5_0, 3 repetitions

	Before	After
Prefill t/s	47.6	51.6 (~+8%)
Decode t/s	24.9	20.7 (no useful gain; noisy / expected for quantized GGUF)

Dense FP16 image gen benefits a lot; ternary/PQ GGUF mostly sees a small prefill bump, as expected.

[Hinoserm]

4x MSI RTX 4070 SUPER Ventus 3X + RTX 4070 Ti + RTX 3090 Ti (VFIO guest) — v2.0.0 (0532473), all unlocked first boot, no Xid

Cards

• 4x MSI GeForce RTX 4070 SUPER Ventus 3X 12 GB — 10de:2783 (subsystem 1462:5137), VBIOS 95.04.69.00.B1, 220 W stock
• 1x MSI GeForce RTX 4070 Ti Ventus 2X 12 GB — 10de:2782 (subsystem 1462:513e), VBIOS 95.04.3C.C0.1A, 285 W stock
• 1x Founders Edition RTX 3090 Ti 24 GB — 10de:2203, VBIOS 94.02.A0.00.01, 450 W stock — passed through to a KVM guest

System

• Ada host: Fedora 44, kernel 7.2.4-200.fc44, NVIDIA 615.71.09 (open), dracut, no Secure Boot. All cards in one server on PCIe 3.0 behind shared switches (4070 Ti at x8).
• 3090 Ti guest: Fedora 44, kernel 7.2.10-200.fc44, NVIDIA 615.71.09 (open), KVM/VFIO. Installed in the guest; activated with a full VM stop/start (virsh shutdown / start). Stayed unlocked across a later host reboot.
• Installed with make install-rpm; gpu_unlock.targets= lists only these cards (the box also has Blackwell 5070s, not targeted).

Benchmark (after unlock)

• Tool: standalone cublasGemmEx (CUDA 13.4), square 8192³, median of 7×20 GEMMs after warmup.
• Stock power limits, one card at a time. This is a production box (shared PCIe 3.0 bus and PSU, cards serving traffic), so this is the cleanest we could do. No "before" GEMM on these exact cards; a locked RTX 5070 from the same box is shown for reference.

Card	FP16 (FP32 acc)	BF16 (FP32 acc)	TF32	FP16 (FP16 acc, ref)
3090 Ti (VFIO guest), 450 W	153.5	157.3	82.1	159.5
4070 Ti, 285 W	153.4	155.9	79.8	146.8
4070 SUPER Ventus 3X, 220 W	134.0	137.2	70.5	117.1
5070 (locked, reference), 250 W	70.6	70.6	34.9	138.9

All figures in TFLOPS.

Real-world (production, before vs after)

• SDXL (UNet denoise, FP16/BF16, no fp8, custom engine), median ms per step-megapixel per card:
  • 4070 SUPERs: 222–231 → 165–176, +26% to +36% throughput (about +31% average).
  • 4070 Ti: 201 → 134–138, +46%.
  • Before = 49–68 production jobs per card; after = production plus a 220-job production-shaped test (220/220 ok, no NaN).
• LLM (Gemma 4 26B-A4B, int4 weights dequantized to BF16, custom engine), 3090 Ti: cold 29.5K-token prompt processing went from 3,400–5,300 to about 6,800 tok/s, about +28%. Decode is bandwidth-bound and unchanged. Temperature-0 output was identical to a pre-unlock output.

Unlock log (4070 SUPER; the others are identical)

gpu_unlock: stage=claim device=0x2783 family=AD102/AD103/AD104/AD106/AD107
gpu_unlock: stage=gfw plm=0xffffff8f ss0=0x20111002 ss1=0x4 wpr2_hi=0x0
gpu_unlock: pass=1/4 ... pass=4/4 attempt=1 ...
gpu_unlock: stage=host ss0=0x88888888 ss1=0x8 plm_feat=0xffffffff plm_wpr_cfg=0xfffff0ff
gpu_unlock: unlocked

[Pepinko81]

RTX 3080 Ti 12 GB (10de:2208) — successful unlock; 200–320 W power-limit sweep

• Card: NVIDIA GeForce RTX 3080 Ti Founders Edition 12GB GDDR6X (GA102, desktop; board vendor not recorded). Single installed GPU, also driving the display.
• OS: Ubuntu 22.04, Linux 6.8.0-136-generic, NVIDIA 550.163.01 (proprietary), Secure Boot off, initramfs-tools.
• Unlock: official gpu-unlock 1.1.0, commit 6081407b312aa3139a2bed62bdda354f4ffddc1d, installed with make install (DKMS); early-boot kernel log: pci 0000:01:00.0: gpu_unlock: unlocked. NVIDIA driver initialized normally afterward.
• Benchmark: PyTorch 2.6.0+cu124, CUDA 12.4, torch.mm(..., out=...), 8192 × 8192 dense GEMM (BF16/FP16 use the respective output dtype; default FP32 internal accumulation), TF32 enabled for the TF32 run; reduced-precision reduction disabled. Five CUDA-event timing samples per point after warm-up. Power limit changed with nvidia-smi; stock clock control (no manual clock locks for the AI tests).

Power cap	BF16 TFLOPS stock → unlocked	FP16 TFLOPS stock → unlocked	TF32 TFLOPS stock → unlocked
200 W	43.8 → 53.3 (+22%)	39.8 → 50.2 (+26%)	28.2 → 30.5 (+8%)
220 W	52.0 → 64.8 (+25%)	49.3 → 60.4 (+23%)	30.0 → 36.9 (+23%)
240 W	57.3 → 75.2 (+31%)	56.1 → 70.7 (+26%)	31.6 → 44.8 (+42%)
260 W	61.2 → 84.9 (+39%)	59.8 → 79.3 (+33%)	33.0 → 50.0 (+51%)
270 W	61.1 → 88.3 (+45%)	59.7 → 82.5 (+38%)	33.0 → 52.3 (+59%)
280 W	63.6 → 92.5 (+46%)	62.2 → 87.9 (+41%)	34.1 → 55.4 (+63%)
300 W	65.4 → 97.8 (+50%)	63.4 → 95.5 (+51%)	34.8 → 57.7 (+66%)
320 W	65.7 → 109.6 (+67%)	65.2 → 103.6 (+59%)	35.6 → 59.2 (+66%)

Controls @320 W: FP32 22.94 → 23.39 TFLOPS (+1.9%); memory copy 808.25 → 805.77 GB/s (−0.3%). BF16 MLP 62.85 → 121.10 TFLOPS (+92.7%); unlocked MLP is already ~121 TFLOPS at 220 W. The 256×256 numerical-check output hashes match before/after for BF16, FP16 and TF32.

Scope/limitations: One independent run per power/phase (five within-run timing samples), so gains are exploratory measurements, not cross-card averages. BEFORE used a modified v3 benchmark and AFTER v5 with equivalent matmul kernels; power-cap values are not measured wall-power savings. GPU-junction memory temperature was unavailable. The 270 W BEFORE run began hotter than AFTER. No long-duration stress or real LLM throughput test was performed.

Related retail 3080 Ti results already exist (e.g. #59, #61, #133, #170); this adds a multi-power sweep on Ubuntu 22.04 / driver 550. Full result tables and anonymized JSON are available separately.


FULL_REPORT.md
benchmark_before_after.csv

[MT-z]

RTX 4090 (AD102) — working unlock, +26% on BF16 image generation

Card: NVIDIA GeForce RTX 4090 (AD102, Palit, PCI 10de:2684)
OS / kernel: Ubuntu, kernel 7.0.0-42-generic
Driver: 615.71.10 (CUDA 13.x)
gpu-unlock: 2.0.0, gpu_unlock.targets=all on the kernel command line
Power limit: 450 W stock (also tested 500 W)

Unlock applies cleanly at boot before the NVIDIA driver binds:

gpu_unlock: stage=host ss0=0x88888888 ss1=0x8 plm_feat=0xffffffff plm_wpr_cfg=0xfffff0ff
gpu_unlock: stage=flr ... ss0=0x88888888 ss1=0x8 plm=0xffffffff ... chip=0x192000a1
gpu_unlock: unlocked

Benchmark — ComfyUI, BF16 DiT image generation

Model: MOSS-Image (DiT in bf16), 1392x1792, 25 steps, CFG 2.5, PyTorch SDPA attention.
Fixed seed, 3 runs each after a warmup. Engine output is bit-identical before/after
(same pixel SHA-256 in every run), so the unlock changes only speed.

	stock (no unlock)	unlocked, 450 W	unlocked, 500 W
s/it	2.153	1.707	1.663
time / image (sampling)	53.7 s	42.4 s	41.5 s
vs stock	—	+26%	+30%
energy / image	22.8 kJ	18.4 kJ	20.0 kJ

Notes:

• The gain is capped by power, not tensor throughput. After unlocking, the card sits on
  the 450 W SW power cap ~95% of the sampling time (stock was limited by the voltage /
  reliability limiter instead), with SM clock ~200 MHz lower.
• Raising the limit to 500 W only adds +2.6% over unlocked-450 W while using ~9% more
  energy per image. 450 W (stock) is the efficiency sweet spot; there's no reason to
  push the power limit for this workload.
• Energy per image actually drops ~19% at stock power — same pictures, less wall-clock,
  less energy.

Where it did and didn't help

• BF16/FP16 matmul-heavy work (image generation): big win (+26%). The DiT forward is
  ~85% tensor-core GEMM (BF16, FP32 accumulate) + SDPA, which is exactly what the unlock
  targets.
• Quantized-GGUF LLM inference (llama.cpp/ggml-based engine): limited. An nsys profile
  of a 41k-token prefill showed only ~25% of kernel time in cuBLAS BF16/FP16 tensor GEMM;
  the bulk is INT8 MMQ expert matmul (~30%) and INT8 attention, which the unlock does not
  touch. Estimated prefill upper bound ~+12%; decode (memory-bound) unchanged.

Overall: clear, stable win for BF16 diffusion/image workloads on a 4090; marginal for
low-bit quantized LLM inference. No instability, no driver changes, survives reboots.

[josh4364]

Worked, thanks.

Claude:

Working on NixOS 🎉

• Card: NVIDIA RTX 3090 Founders Edition 24 GB
• PCI ID: 10de:2204 (subsystem 10de:147d)
• Desktop
• Distribution, kernel and NVIDIA driver version: NixOS 26.05, kernel 6.18.57, NVIDIA 595.71.05 (open kernel module)

Before/after — PyTorch a @ b dense GEMM, 8192×8192, FP32 accumulation (reduced-precision reduction off), stock 350 W power limit:

	Before	After	Gain
FP16	69.1 TFLOPS	108.6 TFLOPS	+57%
BF16	70.0 TFLOPS	112.3 TFLOPS	+60%
TF32	36.3 TFLOPS	61.3 TFLOPS	+69%

NixOS note: gpu_unlock has to be in boot.initrd.kernelModules. A softdep nvidia pre: gpu_unlock modprobe rule didn't work because systemd-modules-load loads nvidia directly and ignores the softdep. Built the module with boot.extraModulePackages and the firmware with hardware.firmware from a flake = false input.

[Esequiel378]

RTX 3090 24 GB: unlocked, +51–61% on real workloads, and an undervolt that only helps once unlocked

• Card: NVIDIA-branded RTX 3090 24 GB, blower cooler, subsystem 10de:1454 (not the FE; likely an OEM card), VBIOS 94.02.42.00.02, stock 350 W
• PCI ID: 10de:2204
• Desktop (headless server)
• Distribution / kernel / driver: NixOS 26.05, kernel 6.18.48, NVIDIA 595.71.05 (open)
• gpu-unlock: 1.0.0 (990a14e), targets=0000:02:00.0 resume=0, loaded from the initrd

Before/after. Three boots (stock → unlocked → stock), median of 10 runs per cell, every output SHA-256-checked against stock. Gains are vs the faster stock session.

	Stock	Unlocked	Unlocked + undervolt
GEMM 8192³ BF16 (FP32 acc.)	71.0 TFLOPS	115.1 (+62%)	121.1 (+71%)
GEMM 8192³ FP16 (FP32 acc.)	69.7 TFLOPS	112.3 (+61%)	118.3 (+70%)
GEMM 8192³ TF32	37.8 TFLOPS	62.8 (+66%)	66.0 (+74%)
ComfyUI Z-Image Turbo BF16, 1024², 8 steps	7.46 s/img	4.95 (+51%)	4.82 (+55%)
ComfyUI SD 1.5 FP16, 1024², 30 steps	6.40 s/img	4.64 (+38%)	4.52 (+41%)
vLLM 0.32 Qwen3-8B BF16 prefill	4,662 tok/s	7,505 (+61%)	7,843 (+68%)
vLLM decode, batch 1	49.4 tok/s	51.1 (+3.5%)	50.6 (+2.4%)

• Outputs: 0 hash mismatches across 260 runs. A 2 h soak (unlocked + undervolt) was 540/540 bit-identical, with no Xid.
• Undervolt: -lgc 210,1695 plus a +150 MHz VF offset cost 3–6% at stock, but added speed once unlocked. The card then sits at its 350 W limit, and less voltage per clock leaves room for more clocks (prefill averaged 1,599 → 1,666 MHz at the same ~335 W).
• Energy: about 40% less per image (Z-Image 2,369 → 1,410 J).

Write-up, raw data, scripts and the NixOS module: https://github.com/Esequiel378/gpu-unlock-3090-results

[jamesmuking5]

gpu-unlock 2.0.0: Colorful RTX 3080 10 GB (GA102, 10de:2206) works

• Card: Colorful GeForce RTX 3080 10 GB (subsystem 7377:1401), VBIOS 94.02.26.48.37, stock card. Power limit raised from 370 W to 400 W and a V/F curve offset applied with LACT (same settings before and after)
• PCI ID: 10de:2206
• Desktop, single GPU that is also the display card (KDE Plasma, Wayland). Secure Boot off
• OS: CachyOS (rolling), kernel 7.3.0-3-cachyos (Clang-built), NVIDIA 615.78.08 open kernel modules (prebuilt linux-cachyos-nvidia-open), mkinitcpio + Limine 12.9.3
• Install: 1fc6d62, make install-arch. DKMS built cleanly for 7.3.0-3-cachyos and 6.18.55-2-cachyos-lts

On CachyOS with Limine, the kernel parameter goes into KERNEL_CMDLINE[default]+="…" in /etc/default/limine, followed by sudo limine-update (rebuilds both initramfs images with zz-gpu-unlock.conf and the boot entries). The first unlocked boot was a one-off edit in the Limine menu (E).

Selectors went from ss0=0x1025055 ss1=0x7 to ss0=0x88888888 ss1=0x8, then unlocked, all stages ret=0. nvidia bound normally after the FLR, no Xid.

PyTorch 2.14.1+cu132 (pytorch/pytorch:2.14.1-cuda13.2-cudnn9-runtime in Docker), torch.matmul, 8192³, FP32 accumulation (allow_tf32=True for TF32), median of 30 after 3 warm-ups, 400 W limit. Before and after are separate boots without and with gpu_unlock.targets=:

Test	Before	After	Change
FP16 GEMM	59.9	116.5	+94%
BF16 GEMM	59.3	118.3	+99%
TF32 GEMM	29.0	57.4	+98%
LFM3-VL-2B BF16 prefill, 1 × 1397 tok (Transformers)	195.6 ms	116.6 ms	−40%
LFM3-VL-2B BF16 prefill, 4 × 1397 tok	7486 tok/s	13028 tok/s	+74%
Peak board power during the run	345 W	401 W	now hits the 400 W cap

GEMM in TFLOPS. Decode stayed at ~45–49 tok/s (memory- and overhead-bound). As a control, the Cyberpunk 2077 built-in benchmark (path tracing / DLSS / native, 1440p) moved only within run-to-run noise. Suspend and numerical parity not tested.

[akeyhero]

RTX 4080 16 GB (AD103, 10de:2704) — FP8 is unlocked too: 203 → 353 TFLOPS (+74%)

• gpu-unlock 2.0.0 (dkms), gpu_unlock.targets=all
• Ubuntu 24.04.5, kernel 7.0.0-38-generic, NVIDIA 615.71.10, VBIOS 95.03.33.00.16
• Stock 320 W power limit, PCIe Gen3 x16 host slot
• Both runs after a full power-off (cold boot)

Dense 8192³ GEMM (FP32 accumulation)	Stock	Unlocked	Gain
FP8 e4m3 (torch._scaled_mm, bf16 out)	202.6	352.5	+74%
BF16	104.9	170.4	+62%
FP16	104.8	148.6	+42%
TF32	52.4	75.0	+43%
FP32 (control)	34.3	33.9	—

Method: torch 2.14.1+cu130, torch.matmul (cuBLAS), reduced-precision reduction off. Median of 20 runs after 3 warm-ups, timed with CUDA events. Outputs are bit-identical before and after: the SHA-256 of a seeded 4096³ product matches for FP32, TF32, FP16 and BF16. Script: https://gist.github.com/akeyhero/cc2e15b384f1b5dd6c2a4759ed6257dc (run with uv run --script bench.py).

End to end: 12-layer d=768 Transformer LM (seq 512, batch 16, BF16 autocast, random tokens)	Stock	Unlocked	Speedup
Training (ms/step)	90.6	67.5	1.34×
Inference (ms/batch)	25.4	18.5	1.37×

My own workload gained much less, because it is not GEMM-bound: training a small ModernBERT (h512, 8 layers) with Hugging Face Accelerate, mixed_precision='fp8' (Transformer Engine backend) and torch.compile, went from 6.51 to 6.87 it/s (+5.5%).

The README lists "INT8, FP8 and INT4 models: No change" and names only FP16/BF16/TF32 as capped. On this Ada card, FP8 GEMM with FP32 accumulation gets unlocked as well: 202.6 TFLOPS stock vs 352.5 unlocked. FP8 can only be measured on Ada and newer, so the "No change" row may come from Ampere reports.

[exyellow9]

Zotac RTX 3060 12 GB (LHR), Ubuntu 26.04, headless desktop: unlocked, ~49 TFLOPS BF16 at stock 170 W, outputs bit-identical

GPU	GeForce RTX 3060 12 GB "Lite Hash Rate" (GA106, 10de:2504), Zotac (subsystem 19da:4630)
System	Desktop, Ryzen 5 5500 / B550, slot 0000:10:00.0, no display attached (headless ComfyUI box)
OS / kernel	Ubuntu 26.04, 7.0.0-38-generic, dracut initramfs, Secure Boot off
NVIDIA driver	595.91.07 open kernel module (unchanged; firmware pin 610.43.03 made no difference)
gpu-unlock	2.0.0 (7eb51bc), make install (deb, DKMS; built cleanly for both installed 7.0.0-31 and 7.0.0-38 kernels, gcc 15.2)
Command line	gpu_unlock.targets=all
Power limit	170 W (stock; the card allows up to 187 W, not tried)
Benchmark	PyTorch 2.14.0+cu130, torch.matmul dense GEMM 8192×8192, 5 runs × 10 iterations after 5 warm-up, PyTorch defaults

Kernel log ends stage=handoff ret=0 ... gpu_unlock: unlocked; nvidia then bound normally, no Xid.

Results (TFLOPS)

	Before	After (median of 5)	Gain
BF16	27.4	49.8 (48.6-50.2)	+82%
FP16	27.4	48.4 (46.7-48.6)	+77%

TF32 and FP32 were not measured.

Real workload: ComfyUI SDXL (perfectionRealisticILXL, 832×1216, 30 steps, dpmpp_2m_sde/karras), warm, fixed seeds: 19.87 s -> 14.17 s per image (1.40x). All three seeds are pixel-identical to the pre-unlock images (sha256 of the decoded PNG pixels).

Video models, INT8 weights — no change, as the README's table says. Same ComfyUI graphs and seeds, exact times from Prompt executed in in the server log:

• LTX 2.5 22B distilled, int8-convrot transformer + int8-convrot Gemma encoder, 57 frames 1344×768, 8 steps, fresh prompt: 120.6-121.6 s before (4 runs) → 116.4 / 117.3 s after.
• MiniMax H3 fl2va int8-convrot + Turbo LoRA, 107 frames 768×1344, 8 steps, 4 reference images, sparse attention: 470 s before → 421 s after, inside the normal shot-to-shot spread.
  (Note for anyone re-timing: ComfyUI caches the text encode by prompt text, so a same-prompt re-run skips the encoder load and looks ~45 s faster for reasons unrelated to the GPU — compare fresh-prompt to fresh-prompt.)

Stability: 180 s sustained BF16 8192² loop held 49.0 TFLOPS flat (48.7 minimum over 18 ten-second windows), 66 °C peak, 1747 MHz SM, throttle reason = software power cap only, 0 Xid.

Notes: this Ubuntu uses dracut (not initramfs-tools); the package's dracut config put the module and firmware into both initramfs images and it worked without changes. Not tested: Wan 2.2, higher power limit.

[Chidwi-commits]

Gainward GeForce RTX 3090 Phoenix 24 GB — 72% higher BF16 throughput at 200 W

On the same card with the power limit set to 200 W for both runs, BF16 throughput increased from 70.14 to 120.59 TFLOP/s after unlocking. Median latency dropped from 15.675 to 9.118 ms.

Environment

• GPU: Gainward GeForce RTX 3090 Phoenix 24 GB (desktop, GA102). PCI ID 10de:2204, subsystem 10b0:1454, VBIOS 94.02.59.00.42.
• Host and guest: Proxmox VE 8.4.1 with VFIO GPU passthrough to a Q35/KVM guest running Ubuntu 24.04.5 LTS, kernel 6.8.0-147-generic, proprietary NVIDIA driver 580.178.04.
• Installation: gpu-unlock 2.0.0 (commit 7eb51bc). After a one-time boot test, I installed the DKMS and firmware packages inside the guest and added gpu_unlock.targets=0000:01:00.0 to its default GRUB entry.

Benchmark

PyTorch 2.9.0+cu128, dense 8192 × 8192 matrix multiplication with torch.mm. I used the same pre-cast BF16 inputs and preallocated output, with FP32 accumulation and reduced-precision reductions disabled. Each result is the median of three CUDA-event timings after one warm-up. All runs passed the correctness check against an FP32 reference with TF32 disabled.

GPU state	Power limit	Median latency	BF16 throughput
Before unlock	200 W	15.675 ms	70.14 TFLOP/s
Unlocked	200 W	9.118 ms	120.59 TFLOP/s
Unlocked	350 W	9.086 ms	121.01 TFLOP/s

Raising the limit from 200 W to 350 W after unlocking changed throughput by 0.35%, within the spread of these short runs.

Guest reboot: The initial unlock logged gpu_unlock: unlocked. On a later warm reboot, the module read SS0=0x88888888 and SS1=0x8 before attempting any writes and returned -EALREADY. The card was already unlocked and reached 120.74 TFLOP/s after reboot.

Stability: The physical Proxmox host rebooted during an earlier 350 W attempt, before that run saved a result. The logs did not establish the cause. The runs in the table completed successfully. Sustained-load stability and end-to-end application performance remain untested.

[p3lusa]

gpu-unlock 2.0.0: RTX 3080 Ti (GA102, 10de:2208) works inside a KVM VM (PCIe passthrough)

• Card: KFA2 RTX 3080 Ti 12 GB, VBIOS 94.02.71.00.28, 350 W default limit, 380 W max
• PCI ID: 10de:2208
• Machine: QEMU/KVM guest with the GPU passed through (QEMU PCI bridge in guest lspci); host Ryzen 5 5600X
• OS: Ubuntu 26.04.1, kernel 7.0.0-38-generic, NVIDIA 595.99.02, DKMS debs
• Secure Boot on: dkms signed the module with the enrolled MOK; only the out-of-tree taint. Unlock verified in early boot log (stage=host ss0=0x88888888 ss1=0x8, unlocked), driver binds normally after.

Packaging note (Ubuntu): make install-deb's apt-get build-dep <local dir> doesn't resolve against a local tree; I ran make deb + apt-get install ./build/deb/*.deb instead, and dh-sequence-dkms is provided by the separate dh-dkms package on Ubuntu 26.04, not by dkms itself.

PyTorch 2.14.0+cu130, 8192³ GEMM, FP32 accumulation (torch defaults), 350 W, stock clocks:

	Before	After	Change
BF16 TFLOPS	73.6	121.0	+64%
FP16 TFLOPS	72.2	115.3	+60%
TF32 TFLOPS	38.8	64.6	+66%
FP16 SDPA TFLOPS	66.4	108.1	+63%
FP32 TFLOPS (control)	25.4	25.5	0%

Sustained behavior (35 s continuous GEMM with live sampling): the card settles at ~350 W / ~1520 MHz at the stock limit, and raising power.limit to the vBIOS max of 380 W only buys +1% (117.6 → 118.7 TFLOPS, ~360 W is its natural equilibrium) — on this card the gain is fully power-bound, matching the pattern others report. Clock locks via nvidia-smi -lgc above the equilibrium clock are a no-op (the cap pulls it back to ~1520 regardless); locking at ~1500 gives stable clocks at 116.0 TFLOPS / 343 W. Zero Xid errors since the unlock; suspend/resume path untested (server VM, never suspends). Passthrough takeaway: the FLR the unlock performs works fine on a fully-assigned device in a guest, and the unlock survives guest reboots without any host-side action.

[snowopsdev]

Note

🤖 Claude Opus 5.5 responding on behalf of snowopsdev

EVGA RTX 3080 12 GB (10de:220a) on Arch/Omarchy with Limine UKI: unlocked, 131.9 TFLOPS BF16 (FP32 accumulate)

• Card: EVGA GeForce RTX 3080 12 GB (GA102, subsystem 3842:4877), VBIOS 94.02.85.40.0A, 400 W default limit
• PCI ID: 10de:220a
• Desktop: compute-only; display runs on an Intel Arc B65, so no blank-screen issue
• OS: Omarchy (Arch), kernel 7.2.15-1-omarchy, nvidia-open-dkms 615.78.08, CUDA 13.4
• gpu-unlock: 2.0.0 (7eb51bc), makepkg from packaging/arch; DKMS also built cleanly for stock 7.2.11-arch1-1
• Command line: gpu_unlock.targets=0000:0e:00.0

Kernel log shows stage=host ss0=0x88888888 ss1=0x8 and ends with gpu_unlock: unlocked; nvidia binds normally afterwards.

Results (cuBLAS cublasGemmEx, 8192³, 5 warm-up + 50 timed iterations, 400 W limit, stock clocks):

	After
BF16, FP32 accumulate	131.9 TFLOPS
FP16, FP32 accumulate	135.7 TFLOPS
TF32	66.4 TFLOPS
FP16, FP16 accumulate (control)	135.2 TFLOPS

No before run. FP32-accumulate throughput now matches the uncapped FP16-accumulate control. The capped ceiling at this card's 2130 MHz max clock is about 76 TFLOPS (70 SMs × 512 FLOP/clk). During the run the card held ~2010 MHz and ~287 W at 41 °C, well under the power limit.

Limine/UKI setup note: the cmdline went into a drop-in, /etc/limine-entry-tool.d/gpu-unlock.conf (KERNEL_CMDLINE[default]+=" gpu_unlock.targets=…"), followed by sudo limine-update. The package's mkinitcpio drop-in needed no edits, and gpu_unlock ran before nvidia bound the card.

Not tested: suspend/resume, sustained load, or real workloads.

[guccukCENEVAR]

Card: MSI GeForce RTX 4060 Ti Ventus 2X Black 16G OC, 16 GB (AD106)
PCI ID: 10de:2803 (16 GB board reporting 2803, 16380 MiB per nvidia-smi)
Desktop
Distro / kernel / driver: CachyOS, 7.3.1-1-cachyos, NVIDIA 615.78.08 (CUDA 13.4)

Used the Ada fork JustAResearcher/gpu-unlock-ada 1.1.1, with gpu_unlock.targets=<exact address> gpu_unlock.ad106_experimental=1, cold boot. VBIOS 95.06.34.00.F3.

No raw FP16/BF16/TF32 TFLOPS numbers yet. Real workload, same card, same 165 W power limit, both states:

llama.cpp (llama-bench, Qwen3-4B BF16, -ngl 99):

	stock	unlocked	change
pp512	4286 t/s	5585 t/s	+30.3%
pp2048	4225 t/s	5489 t/s	+29.9%
tg128	31.1 t/s	31.3 t/s	~0 (memory-bound)

30 min sustained pp2048: 4188 → 5415 t/s median (+29.3%). Same power (~152–154 W, at the cap ~92% of the time in both states), same max temp (78 °C), no thermal throttling, no Xid errors.

[mikecavallo]

2× RTX 3090 24 GB (ASUS + MSI, 10de:2204) on Ubuntu 24.04: both unlocked, ~116 TFLOPS FP16 / ~119 BF16 at 350 W

• Cards: ASUS RTX 3090 24 GB (subsystem 1043:87af, VBIOS 94.02.42.00.A9) and MSI RTX 3090 24 GB (subsystem 1462:3884, VBIOS 94.02.42.00.F1)
• PCI ID: 10de:2204 (both)
• Desktop: i9-14900KS, 1350 W PSU. The ASUS card (01:00.0) also drives the display.
• OS: Ubuntu 24.04.5 LTS, kernel 7.0.0-42-generic, proprietary NVIDIA driver 580.178.04, Secure Boot off
• gpu-unlock: 2.0.0 (7eb51bc), make install (deb, DKMS)
• Command line: gpu_unlock.targets=all

The kernel log shows stage=host ss0=0x88888888 ss1=0x8 and gpu_unlock: unlocked for both cards. nvidia binds normally afterwards.

Results: PyTorch 2.9.0+cu128, torch.matmul 8192³, FP32 accumulation (reduced-precision reductions off), 10 warm-up + 300 timed iterations (100 for the 250 W rows), CUDA-event timing. Unlocked in both rows; no before run.

	Power limit	FP16	BF16	TF32
ASUS (01:00.0, display)	250 W	68.0	73.3	43.5
MSI (02:00.0)	250 W	79.3	82.0	50.0
ASUS (01:00.0, display)	350 W	104.0	107.1	58.1
MSI (02:00.0)	350 W	115.5	118.5	64.1

TFLOPS. Under load at 350 W the cards ran at ~1425–1600 MHz SM, which works out to ~870–970 FLOP/clk/SM for FP16/BF16 and ~440–500 for TF32. That's roughly 2× the capped rate (512 / 256). At 250 W the clocks dropped to ~800–1200 MHz, but FLOPs per clock stayed the same, so the unlock holds at both limits.

Install note: make install failed when the checkout was in a path containing a space (~/Github Repos/gpu-unlock). Several Makefile recipes use $(CURDIR) / $(abspath ...) unquoted, so apt-get build-dep and rm -rf got the path split in two. Cloning to a path without spaces fixed it.

Not tested: suspend/resume, sustained load, real workloads.

[felixilars]

RTX 4080 Laptop GPU (Manual PCI ID Patch)

• Card: Dell NVIDIA GeForce RTX 4080 Laptop GPU 12 GB (AD104M, GN21-X9, max 175W TGP)
• PCI ID: 10de:27e0
• Form Factor: Laptop
• OS / Kernel / Driver: Pop!_OS 22.04 LTS | Kernel 7.1.1-76070101-generic | NVIDIA Driver 580.178.04

Setup Notes (Mobile Tweaks)

1. Added PCI ID: 0x27e0 was not in upstream recipes.c, so I manually added AD104(0x27e0), to the AD104 table and rebuilt via DKMS.
2. Mobile D3cold Fix: On hybrid laptops, power management places the dGPU into D3cold on idle, which resets the volatile SS0 registers back to stock 1/2 speed. Fixed by setting options nvidia NVreg_DynamicPowerManagement=0x00 in /etc/modprobe.d/ and forcing dedicated graphics mode.

1. Synthetic GEMM & Attention (PyTorch, FP32 Accumulate)

Test / Precision	Matrix / Context Size	Stock (default)	Unlocked	Speedup
BF16 GEMM	4096 × 4096	60.5 TFLOPS	122.1 TFLOPS	+102.0% (2.02x)
BF16 GEMM	8192 × 8192	67.6 TFLOPS	112.0 TFLOPS	+65.5% (power-capped ~175W)
FP16 GEMM	8192 × 8192	61.8 TFLOPS	106.5 TFLOPS	+72.3%
TF32 GEMM	8192 × 8192	31.3 TFLOPS	53.1 TFLOPS	+69.9%
FP32 CUDA (Control)	8192 × 8192	20.0 TFLOPS	20.2 TFLOPS	~0% (Tensor-only effect confirmed)
FlashAttention (SDPA)	2048 context (BF16)	1.39 ms	0.86 ms	1.62x faster (-38% latency)
FlashAttention (SDPA)	8192 context (BF16)	18.95 ms	11.23 ms	1.69x faster (-41% latency)

2. Real-World Local LLM Inference (Qwen 2.5 7B)

• Ollama (qwen2.5:7b Q4_K_M GGUF):
  • Generation (Decode): 79.1 tok/s $\rightarrow$ 88.0 tok/s (+11.3%)
  • Prompt Prefill: 177,462 tok/s $\rightarrow$ 191,031 tok/s (+7.6%)
• vLLM (Qwen2.5-7B-Instruct-AWQ 4-bit):
  • Throughput: 2,726.0 tok/s $\rightarrow$ 2,767.7 tok/s (+1.5%)

[YunagiHisa]

two RTX 3060 12 GB GPUs work under Proxmox passthrough

• Cards: desktop, 2 × RTX 3060 12 GB, GA106, PCI ID 10de:2504;
  GPU 0 is ZOTAC, second board vendor not independently established.
• Setup: Ryzen 9 5900X, Proxmox PCI passthrough, Ubuntu 24.04 guest,
  kernel 6.8.0-147-generic, NVIDIA 580.178.04, 170 W per GPU.
• Unlock: 2.0.0 at 1fc6d628c49c7fba34ef3a2e7f758700d35ce890.
  Both guest devices (0000:00:10.0, 0000:00:11.0) logged
  gpu_unlock: unlocked, with successful handoffs and normal NVIDIA recognition.
• CUDA: driver compatibility display 13.0; PyTorch 2.11.0+cu128 with
  runtime 12.8; saved llama.cpp CMake records identify build Toolkit 12.8.61.
  llama.cpp build 11669, commit 620d2225c.

BF16 8192³ GEMM, FP32-accumulation policy, means of three independent processes:

GPU	Before	After	Reverted	After / Before
GPU 0	26.825	48.312	26.781	+80.10%
GPU 1	26.924	47.802	26.925	+77.54%

Units: TFLOPS. Across all tested sizes, FP16 improved 62.45–78.48%, BF16
65.22–80.75%, and TF32 70.27–79.10%. FP32 IEEE changes were very small and
negative: −0.24% to −0.01%. Qwen3-8B BF16 pp512 improved 16.31%; Q4 prefill
and generation changed little. Host shutdown and operator-performed AC power
removal preceded After and Reverted; Reverted returned near Before.

Thanks for GPU-Unlock. These are diagnostic measurements from one host and
two cards, not a safety guarantee or proof of complete internal-state restoration.
Full methods, all configurations, SDs, telemetry and limitations:
full report.

[pedrox86lopes]

Card: Acer Nitro V15 (ANV15-51-58AZ), NVIDIA GeForce RTX 3050 6 GB Laptop GPU (GA107BM)
PCI ID: 10de:25ac
Laptop or desktop: Laptop
Distribution, kernel and driver: Omarchy (Arch Linux), kernel 7.2.8-9-omarchy, NVIDIA 615.78.08

Before/after numbers:

• Tool: PyTorch torch.matmul
• Matrix size: 4096×4096 (FP16)
• Power limit: 50 W
• Before (stock): 12.71 TFLOPS (10.81 ms)
• After (unlocked): 18.04 TFLOPS (7.62 ms) — +41.9% speedup

Notes:

• Card required adding GA107(0x25ac) with chip ID 0x177 mapped to &ga10x in recipes.c.
• Boot parameters used: gpu_unlock.targets=all nvidia.NVreg_DynamicPowerManagement=0x00 (disabling D3cold runtime PM is required so Optimus does not revert the unlock state).

[kadiremrah]

Result

Successfully unlocked the NVIDIA GeForce RTX 4060 Laptop GPU (AD107) using gpu-unlock 2.0.0 on CachyOS.

The kernel log confirmed successful completion of the unlock sequence, including all four passes returning ret=0, followed by:

gpu_unlock: unlocked

To verify the performance impact, I performed an A/B/A comparison on the same GPU:

1. GPU Unlock enabled.

2. GPU Unlock disabled for one boot by removing gpu_unlock.targets=all from the GRUB kernel command line.

3. GPU Unlock enabled again after a normal reboot.

The benchmark results show substantial and repeatable improvements in Tensor Core accelerated matrix multiplication.

Card

NVIDIA GeForce RTX 4060 Laptop GPU, 8 GB VRAM.

BIOS-configured power limit: 70 W.

The NVIDIA driver reports a maximum configurable power limit of 105 W, but this was not changed during testing.

PCI ID

device=0x28a0
family=AD102/AD103/AD104/AD106/AD107

Distribution

CachyOS (Arch Linux-based), Fish shell.

Kernel

7.3.1-1-cachyos

NVIDIA driver

615.78.08

CUDA UMD version: 13.4.

Benchmark environment

PyTorch: 2.11.0+cu128
CUDA runtime: 12.8
NumPy: 2.5.4
GPU: NVIDIA GeForce RTX 4060 Laptop GPU

Benchmark configuration:

• Matrix multiplication: 4096 × 4096

• Operation: torch.mm

• Warmup: 20 iterations

• Measurement: 60 iterations

• CUDA event timing

• FP32, TF32, FP16 and BF16

• Three repeated benchmark runs for each comparison state

Performance results

All values are in TFLOPS, averaged over three benchmark runs.

Precision | Locked | Unlocked | Improvement -- | -- | -- | -- FP32 | 6.96 | 6.90 | −0.9% TF32 | 13.61 | 20.12 | +47.8% FP16 | 27.77 | 42.53 | +53.2% BF16 | 28.27 | 51.33 | +81.6%

Individual BF16 measurements

Locked:

28.26 TFLOPS
28.26 TFLOPS
28.28 TFLOPS

Unlocked:

54.34 TFLOPS
46.84 TFLOPS
52.81 TFLOPS

BF16 performance showed greater variation in the unlocked state, but every unlocked measurement substantially exceeded the locked baseline.

Verification of the locked state

For the locked-state benchmark, I temporarily removed the following kernel parameter from the GRUB boot entry:

gpu_unlock.targets=all

After rebooting, /proc/cmdline confirmed that the parameter was absent.

The gpu_unlock kernel module was still loaded, but the kernel log contained no unlock execution or completion messages.

The module's mkinitcpio configuration explicitly states that it performs no unlock operation unless gpu_unlock.targets= is present on the kernel command line.

This allowed comparison without uninstalling the module, changing the NVIDIA driver, or modifying the BIOS power limit.

GPU power and temperature

During the unlocked benchmark, NVIDIA telemetry showed:

GPU utilization: 100%
Peak power draw: 70.47 W
SM clock: approximately 1770–1905 MHz
Memory clock: up to 8001 MHz
GPU temperature: approximately 55–57°C

The 70 W power limit was configured in the laptop BIOS and intentionally left unchanged.

Interpretation

The FP32 results remained essentially unchanged, while TF32, FP16 and BF16 matrix multiplication improved significantly.

This pattern is consistent with the modification affecting Tensor Core accelerated computation rather than general-purpose FP32 throughput.

The A/B/A testing procedure provides strong evidence that the performance difference is associated with GPU Unlock.

These results apply to dense matrix multiplication. No LLM inference, quantized model serving or real-world application benchmarks were performed.

Installation

Installed gpu-unlock 2.0.0 using the DKMS and firmware packages, with the module included in the mkinitcpio-generated initramfs.

The following kernel parameter was added to GRUB:

gpu_unlock.targets=all

The module successfully loaded before the NVIDIA driver initialized the GPU.

Conclusion

gpu-unlock 2.0.0 works on the RTX 4060 Laptop GPU (AD107, device 0x28a0).

On this system, it produced approximately +48% TF32, +53% FP16 and +82% BF16 throughput in the tested dense GEMM workload, with essentially unchanged FP32 performance.

Thanks to the developers for making this modification available.

Did gpu-unlock work on your card? Post it here as a comment instead of opening an issue.

Please include:

* **Card**: brand, model and memory size, e.g. EVGA RTX 3090 FTW3 24 GB

* **PCI ID**: from `lspci -nn -d 10de:`, e.g. `10de:2204`

* **Laptop or desktop**

* **Distribution, kernel and NVIDIA driver version**

* **Before/after numbers** if you have them: tool, matrix size, accumulation type, power limit, and TFLOPS for FP16, BF16 and TF32. Real workloads (LLM prefill, training, image generation) are welcome too.

Keep it short. Logs aren't needed when it worked.

If the unlock failed, gave no speedup, or stopped working later (after idle, suspend or a driver reload), open an unlock failure issue instead.

Earlier success reports were filed as issues and have been closed in favour of this thread. They are still searchable under closed issues.

[Patrity]

Ran this against the stuff I actually use, and holy crap, it's real.

• Cards: ASUS Strix, Zotac and PNY RTX 3090 24 GB (10de:2204), desktop, no PCIe switch. A fourth 3090 is unlocked too, but I didn't benchmark it on its own.
• System: Ubuntu 24.04, kernel 6.8.0-124, driver 580.173.02 (open), gpu-unlock 2.0.0 (7eb51bc). 0 Xid across the run.

8192³ GEMM at 350 W, TFLOPS:

Card	BF16	FP16	TF32	FP32 control
Strix	72.0 → 117.0 (+63%)	70.7 → 111.9 (+58%)	38.3 → 63.1 (+65%)	24.8 → 24.8
Zotac	70.6 → 114.4 (+62%)	69.4 → 111.0 (+60%)	35.2 → 62.4 (+77%)	20.9 → 22.1 (+6%)
PNY	70.0 → 111.1 (+59%)	68.2 → 107.5 (+58%)	37.0 → 60.6 (+64%)	23.8 → 23.8

An RTX PRO 6000 I left locked stayed within 2% on every metric, so I believe the rest.

• Image and video: Qwen-Image 2.1 BF16 +57%, Wan 2.2 5B FP16 +46%
• Search: bge-m3 +40%, mxbai-rerank +43%
• LLM prefill: Qwen3.5-9B BF16 on vLLM +54% at 8K, Qwen3-VL-8B INT4 on vLLM +52%
• Quantized GGUF: between -1% and +5%, as advertised

Full breakdown, including where I land against the README table: #301

Write-up: https://www.techhivelabs.net/blog/gpu-unlock-rtx-3090-benchmark
All 370 comparisons: https://www.techhivelabs.net/data/blog/gpu-unlock/summary.csv

Thanks for building this. It's staying on.

[dinukasal]

Worked on my RTX 2060 Super 8GB VRAM Nvidia graphics card, prefill speed increased from around 120tok/s to 400tok/s when running qwen3.8-flash-next-coder with Strata (https://github.com/mach10x/strata) and Decode increased from ~20tok/s to ~30-35tok/s

[srfrnk]

Hardware & Setup

• Card: EVGA GeForce RTX 3090 FTW3 Ultra 24 GB (subsystem 3842:3975)
• VBIOS version: 94.02.42.80.C1
• Form factor: Desktop (single GPU driving active physical monitors)
• Motherboard & CPU: Gigabyte B460M D3H, Intel Core i9-10900 (Comet Lake)
• OS & Kernel: Ubuntu 24.04 LTS, kernel 7.0.0-42-generic / 7.0.0-38-generic
• NVIDIA Driver: 595.99.02 (proprietary)
• gpu-unlock version tested: Release 2.0.0 (7eb51bc)

Context & Inspiration

I saw the success report in #132 by @stefb69 on the exact same card (3842:3975), identical VBIOS (94.02.42.80.C1), and same 10th-Gen Intel architecture
(i7-10700). However, attempting the unlock on my setup currently encounters a freeze during the early PCIe reset phase, and I would welcome ideas or suggestions
from anyone who might know how to work around this on similar platforms.

Current Observations & Trials

1. Early Boot Initramfs (Dracut & initramfs-tools):
   • Configured force_drivers+=" gpu_unlock " with the priority 00-gpu-unlock.conf prefix to guarantee gpu_unlock loads before nvidia.
   • Booting with gpu_unlock.targets=0000:01:00.0 halts with a blank screen / freeze during early boot right when the PCIe FLR reset takes place on the primary
     display card.
2. Motherboard iGPU Attempt (IGFX):
   • The i9-10900 has integrated Intel UHD 630 graphics, which could theoretically drive display while leaving the 3090 headless for the unlock. However, selecting
     Initial Display Output = IGFX in this Gigabyte B460M BIOS while a PEG card is seated resulted in no video output across all ports (had to clear CMOS to
     restore).
3. Pre-Boot UEFI Direct MMIO Probe:
   • Tested direct register reads/writes via a standalone EFI probe tool (0x0082381c SS0 / 0x00823820 SS1). As expected, the registers remain hardware-locked at
     0x00000000, confirming the authenticated Falcon SEC2 microcode + FLR sequence is required.

Question for the Community

Has anyone with a Gigabyte B460 / 400-series Intel motherboard (or single primary display card on Ubuntu) successfully navigated the early pcie_flr framebuffer
handoff?

• Are there specific kernel parameters (nvidia-drm.modeset, video=efifb:off, pci=..., etc.) or BIOS settings that help prevent the PCIe root port from hanging
  during early reset when driving the active display?

Any insights, debugging suggestions, or workarounds would be greatly appreciated!

[timofey]

RTX 4090 24 GB (Dell subsystem 1028:5107): unlock works, up to +51% in matmul benchmarks

• PCI ID: 10de:2684, subsystem 1028:5107
• VBIOS: 95.02.3C.80.E7
• Setup: single card, also the display card, 450 W stock power limit
• Driver: 615.78.08, open kernel modules (nvidia-open)
• OS: Arch Linux, kernel 7.2.11-arch1-1, mkinitcpio
• gpu-unlock: 2.0.0 (gpu-unlock-dkms 2.0.0-1)

Log: stage=flr ... ss0=0x88888888 ss1=0x8, stage=handoff ret=0 ... cleanup=0, gpu_unlock: unlocked

Square torch.mm with PyTorch 2.10 (CUDA 13.0), FP32 accumulation. 30 s heat-up, then 2 s warmup and 6 s measured per test. Median TFLOPS, one run per state:

Test	Stock	Unlocked	Gain	SM clock, stock → unlocked
TF32, n=8192	85.2	128.3	+51%	2632 → 1995 MHz
BF16, n=2048	163.7	234.1	+43%	2626 → 2085 MHz
FP16, n=2048	152.2	176.2	+16%	2515 → 2325 MHz
FP16, n=8192	153.8	165.7	+8%	2378 → 2262 MHz
BF16, n=4096	145.2	146.6	+1%	2518 → 2520 MHz
BF16, n=8192	140.9	142.3	+1%	2438 → 2428 MHz
FP16, n=4096	142.9	143.5	+0.4%	2429 → 2416 MHz

Nearly every test sat at the 450 W limit in both states, so large BF16/FP16 matmuls gain little here, as the README warns for the 4090. The gain varies a lot with matrix size.

[Manos1966]

Card: NVIDIA GeForce RTX 3060, 12 GiB VRAM
PCI ID: from lspci -nn -d 10de:, e.g. 10de:2204
Server: 2 x Intel Xeon E5-2690 v2 @ 3.00GHz (20 cores, 40 threads); installed RAM: 310.6 GiB
Distribution: Linux 7.0.0-42-generic; GPU driver: 615.71.09; Strata: project 0.1.51, commit 4b9e05a2; engine 0.1.51
Before/after numbers: did not have time to test
It worked! Torch test scores 48,4 TFLOPS

a big THANKS!