[Results] gpu-unlock on 3x RTX 3090: 30+ real model setups before and after (image, video, LLM, embeddings, speech)

[Patrity]

I own four RTX 3090s and they do everything around here: images, video, embeddings, search, coding agents. The matrix benchmarks for gpu-unlock already looked great, so I wanted to know what it does to the stuff I actually run. (I also wanted to catch myself being wrong, because a 60% number deserves suspicion.)

So I ran more than 30 model and format setups, stock and unlocked, on one rig. The short version: the category barely predicts the gain, and the math path predicts almost all of it.

• Story and charts: https://www.techhivelabs.net/blog/gpu-unlock-rtx-3090-benchmark
• All 370 comparisons: https://www.techhivelabs.net/data/blog/gpu-unlock/summary.csv
• Raw results: https://www.techhivelabs.net/data/blog/gpu-unlock/results.jsonl

The Rig

• Cards: ASUS Strix, Zotac and PNY RTX 3090 24 GB (10de:2204). A fourth 3090 is unlocked too, but I didn't benchmark it on its own.
• System: EPYC 7532, 128 GB, Ubuntu 24.04, kernel 6.8.0-124, driver 580.173.02 (open), no PCIe switch
• gpu-unlock: 2.0.0 at 7eb51bc, targeted by PCI address
• Software: ComfyUI, vLLM, llama.cpp, transformers, sentence-transformers
• Stability: 0 Xid across the whole run. All four have stayed unlocked since.

How I Tried to Fool Myself

• Same everything: same seeds, same prompts, same 350 W limit, all AI services stopped for both phases
• Locked control: an RTX PRO 6000 I never unlocked stayed within 2% on every metric
• FP32 control: flat on Strix and PNY. Zotac moved +6% here and -3% in the power sweep. I read that as noise, but I haven't proven it.
• Solo reruns: the cards run side by side, so I repeated a batch with Strix alone. Same numbers.
• Outputs: same-seed results matched in 75 of 81 test groups. Z-Image on Strix wouldn't even reproduce against itself within one phase.

I also threw out my 4-card Z-Image throughput test. That one's on me: a metric bug let the slowest card set the score, and I overwrote the stock logs before I caught it.

GEMM

8192³ 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

Real Workloads

One card each, 350 W. LLM rows are at an 8K prompt.

Workload	Format	Card	Stock	Unlocked	Gain
Qwen-Image 2.1	BF16	Strix	26.15 s/image	16.60 s/image	+57%
FLUX.3 Klein	BF16	Strix	10.14 s/image	6.48 s/image	+57%
Z-Image Turbo	BF16	Strix	7.70 s/image	5.06 s/image	+52%
SDXL base	FP16	Strix	8.51 s/image	6.28 s/image	+35%
Wan 2.2 5B	FP16	Strix	107.1 s/video	73.4 s/video	+46%
Qwen3-Embedding-4B	FP16	Strix	7,352 tok/s	10,384 tok/s	+41%
bge-m3	FP16	Strix	81,348 tok/s	113,854 tok/s	+40%
mxbai-rerank-large-v2	FP16	PNY	43.4 pairs/s	62.1 pairs/s	+43%
Qwen3-Reranker-4B	BF16	Zotac	14.7 pairs/s	20.7 pairs/s	+41%
Whisper large-v3-turbo	FP16	Zotac	83x realtime	93x realtime	+12%
Qwen3.5-9B prefill, vLLM	BF16	Strix	3,988 tok/s	6,139 tok/s	+54%
Gemma 4 E4B prefill, vLLM	BF16	Strix	6,297 tok/s	8,696 tok/s	+38%
Qwen3-VL-8B prefill, vLLM	INT4	PNY	3,584 tok/s	5,459 tok/s	+52%
Ornith-1.5-9B prefill, llama.cpp	BF16 GGUF	Zotac	2,602 tok/s	3,507 tok/s	+35%
Ornith-1.5-9B prefill, llama.cpp	Q4 GGUF	Zotac	3,201 tok/s	3,260 tok/s	+2%

Decode gained about 4 to 5% on vLLM at 8K and stayed flat on llama.cpp. Exactly what your table says it should do.

Where I Land Against the README Table

Most of it lines up. These are the spots worth a second look, offered as data points since I only have my one rig:

• Wan 2.2: I got +46% on the 5B FP16 model against +15 to 20% in the table. I don't know which variant the table means, so these may not actually conflict.
• INT4: Qwen3-VL-8B on vLLM gained +44 to 52% on prefill and +2 to 8% on decode. That fits your AWQ/GPTQ row, but it reads oddly next to "INT4: no change".
• FP8: Krea 2 Turbo ships as FP8 weights and gained +50%. Your image row does list Krea, while the no change row lists FP8. ComfyUI on my rig logs BF16 and FP16 compute, which would explain it, but I never tied that to the Krea run itself.
• INT8: small, and not quite flat. LTX-2.5 +11%, Qwen-Image +9%, MiniMax H3 +8%, FLUX.3 Klein -4%.
• Quantized GGUF: every Q4 and Q8 run stayed between -1% and +5%, across LLMs, speech, vision, and a decision model.

Power

• Energy: the cards pull about the same watts and finish sooner, so Qwen-Image used 37% less energy per image on Strix. Whisper went the wrong way (+22% energy) and I'd love for someone to tell me why.
• Clocks: on the sustained run, Strix held 1,644 MHz stock and 1,347 MHz unlocked at the same 347 W per nvidia-smi. That's my best read on why it isn't a full 2x.
• Power caps: the unlock needs headroom to spend.

Limit	Strix BF16	Zotac BF16	PNY BF16
200 W	+21%	+27%	+22%
250 W	+25%	+34%	+22%
300 W	+48%	+56%	+41%
350 W	+61%	+61%	+60%

Strix at 390 W added another 4 to 7% on compute bound work and under 1% on decode.

The Parts I Can't Explain

• Splitting across cards: Gemma 4-12B across Strix and Zotac gained +17% on prefill, and a 2-card SDXL CFG split gained +6%. Each card alone gets a much bigger lift, and I haven't isolated what's eating it.
• Qwen3-ASR: it's BF16 and only gained 2 to 4%. No idea.

If your card disagrees with me, I want to hear about it. Thanks for building this. I'd feel worse about those half-rate years if I weren't rendering images 57% faster out of spite.

[misakaai-dev]

Appreciate you calling out that the fourth card was unlocked but outside the benchmark. It makes the three-card results easier to interpret.