NVIDIA RTX PRO 4000 Blackwell, AI & Machine Learning Benchmarks & Specs

24GB · AI Score 7.6/100 · first-party measured on 12 AI workloads

7.6 AI Score ✓ Measured

Every number on this page is first-party: NVIDIA RTX PRO 4000 Blackwell was run on our pinned 12-workload AI suite on 2026-07-12, with under 0.5% run-to-run variance. On Llama 3.1 8B (Q4_K_M) NVIDIA RTX PRO 4000 Blackwell delivers about 112.36 tokens/sec. Stepping up to Qwen3 32B it holds roughly 27.8 tok/s. Llama 3.3 70B does not fit. It needs roughly 42GB and this card has 24GB. For image generation, SDXL runs at 4.28 it/s, while FLUX.1-dev won't fit at BF16 (needs ~26GB). 4 of the 12 workloads won't fit on 24GB at the tested precision, Llama 3.3 70B, FLUX.1-dev, FLUX.1 Kontext, Qwen-Image-Edit. We publish those as hard gates rather than quietly dropping to a smaller quant. NVIDIA RTX PRO 4000 Blackwell isn't a retail purchase for most people. It's rented by the hour. You can run this exact card on RunPod.

AI & Machine Learning benchmark results

Text Generation tok/s 18

MiniCPM5 2B271.18

LFM2.5 2.6B254.94

Qwen3 30B A3B185.31

Granite 4.1 3B182.59

Qwen3 4B179.58

Nemotron 3 Nano 4B174.55

Gemma 4 26B A4B134.51

DeepSeek Coder 7B Instruct v1.5117.35

Llama 3.1 8B112.36

Llama 3 8B104.38

Qwen3 8B101.59

Ornith 1.5 9B91.88

Workload	Result	Telemetry	Data
MiniCPM5 2B	271.18 tok/s	81 W46°CQ4_K_M	✓ Measured
LFM2.5 2.6B	254.94 tok/s	83 W45°CQ4_K_M	✓ Measured
Granite 4.1 3B	182.59 tok/s	90 W52°CQ4_K_M	✓ Measured
Nemotron 3 Nano 4B	174.55 tok/s	90 W48°CQ4_K_M	✓ Measured
Qwen3 4B	179.58 tok/s	2.7 GB peak96 W43°C1.88 tok/WQ4_K_M	✓ Measured
DeepSeek Coder 7B Instruct v1.5	117.35 tok/s	106 W53°CQ4_K_M	✓ Measured
Llama 3 8B	104.38 tok/s	103 W53°CQ4_K_M	✓ Measured
Llama 3.1 8B	112.36 tok/s	4.9 GB peak112 W47°C1.01 tok/WQ4_K_M	✓ Measured
Qwen3 8B	101.59 tok/s	107 W53°CQ4_K_M	✓ Measured
Nemotron Nano 9B v2	82.32 tok/s	109 W53°CQ4_K_M	✓ Measured
Ornith 1.5 9B	91.88 tok/s	105 W51°CQ4_K_M	✓ Measured
Gemma 4 12B	64.44 tok/s	108 W53°CQ4_K_M	✓ Measured
Qwen2.5-Coder 14B	60.62 tok/s	8.6 GB peak123 W53°C0.49 tok/WQ4_K_M	✓ Measured
Qwen3 14B	56.63 tok/s	114 W54°CQ4_K_M	✓ Measured
Gemma 4 26B A4B	134.51 tok/s	84 W50°CQ4_K_M	✓ Measured
Qwen3 30B A3B	185.31 tok/s	74 W50°CQ4_K_M	✓ Measured
Qwen3 32B	27.8 tok/s	18.7 GB peak114 W56°C0.24 tok/WQ4_K_M	✓ Measured
Llama 3.3 70B	✕ Won't fit needs ~46 GB	VRAM-gated at this precision	✓ Measured

Image Generation images/min 7

Stable Diffusion 1.540.91

Sana 1.6B20.12

PixArt-Sigma XL13.44

Stable Diffusion XL8.56

Playground v2.55.54

Z-Image Turbo4.425

Workload	Result	Telemetry	Data
Stable Diffusion 1.5	40.91 images/min	140 W51°C	✓ Measured
Sana 1.6B	20.12 images/min	145 W60°C	✓ Measured
Stable Diffusion XL	8.56 images/min	14.4 GB peak145 W61°C7 s/img	✓ Measured
Playground v2.5	5.54 images/min	145 W72°C	✓ Measured
PixArt-Sigma XL	13.44 images/min	145 W65°C	✓ Measured
Z-Image Turbo	4.43 images/min	23.1 GB peak145 W64°C13.6 s/img	✓ Measured
FLUX.1 dev	✕ Won't fit needs ~26 GB	VRAM-gated at this precision	✓ Measured

Image Editing images/min 2

Workload	Result	Telemetry	Data
FLUX.1 Kontext dev	✕ Won't fit needs ~26 GB	VRAM-gated at this precision	✓ Measured
Qwen-Image-Edit	✕ Won't fit needs ~42 GB	VRAM-gated at this precision	✓ Measured

Video Generation frames/s 2

Workload	Result	Telemetry	Data
LTX-Video (distilled)	3.1 frames/s	9.3 GB peak118 W62°C31.3 s/clip	✓ Measured CPU offload
Wan 2.2 5B (720p)	0.26 frames/s	17.3 GB peak139 W67°C188.5 s/clip	✓ Measured CPU offload

How we measured this. Every result comes from our own pinned, reproducible AI suite, 12 workloads: the Qwen3-4B to Llama-70B LLM ladder (llama.cpp, Q4_K_M), SDXL / Z-Image / FLUX-dev generation, FLUX-Kontext / Qwen-Edit editing, and LTX / Wan video, run first-party on rented hardware with under 0.5% run-to-run variance. Peak VRAM, power draw, temperature and tokens-per-watt are captured per workload. “Won’t fit” rows are real data: where a model exceeds the card’s VRAM at the tested precision we record a hard gate rather than silently dropping to a smaller quant. Measured 2026-07-12 · harness 2.0.0-standalone.

NVIDIA RTX PRO 4000 Blackwell specifications

Architecture	Blackwell
CUDA cores	8,960
VRAM	24GB GDDR7 ECC
Memory bus	192-bit
Memory bandwidth	672 GB/s
Boost clock	2,617 MHz
TDP	140 W
Process	4nm (TSMC 4N)
Interface	PCIe 5.0 x16
Release date	2025-08-01
Launch MSRP	$1,500

Verdict, capable, but 24GB sets the ceiling

NVIDIA RTX PRO 4000 Blackwell scores 7.6/100, #34 of 102. It ran 8 of 12; 4 exceeded its 24GB. Every figure here is our own measurement.

Relative performance: where the NVIDIA RTX PRO 4000 Blackwell lands

100% = this card, AI & Machine Learning headline metric (AI Score). #8 of 20 workstation cards in this vertical.

GPU	%	AI Score
NVIDIA RTX A6000	276%	21
NVIDIA RTX PRO 4500 Blackwell	170%	12.9
NVIDIA RTX A5500	111%	8.4
NVIDIA Quadro RTX 8000	108%	8.2
NVIDIA RTX PRO 4000 Blackwell	100%	7.6
NVIDIA Quadro RTX 6000 (Turing)	97%	7.4
NVIDIA RTX A5000	97%	7.4
NVIDIA RTX 4500 Ada Generation	78%	5.9
NVIDIA RTX A4500	67%	5.1

One tier up: NVIDIA Quadro RTX 8000 (+8%) · compare head-to-head
One tier down: NVIDIA Quadro RTX 6000 (Turing) (-3%) · compare head-to-head

← All AI & Machine Learning GPU rankings

The silicon

Transistors	45,600 million
Die size	378 mm²
Process node	4 nm
Fabricated by	TSMC
Transistor density	120.6 million per mm²

Denser than 92% of the 746 cards we have silicon data for. Density is the clearest measure of what a process node bought: a card that gained it without growing the die got its speed from the fab rather than the architecture.

Silicon figures from Wikipedia (CC BY-SA 4.0). Benchmarks on this page are our own. Compare every chip.

What this card can build

Whole-job timings, composed from our measured per-model results on this card.

Workflow	Time	Energy	Basis
24-frame storyboard	6.4 min	14.71 Wh	all 2 stages measured
60-second AI short film	10.2 min	20.63 Wh	all 3 stages measured
Full codebase review	16.5 min	33.79 Wh	measured
Short social clips	34.2 min	78.69 Wh	all 3 stages measured

Can't run: Product photo shoot (needs FLUX.1 Kontext dev), Photo restoration batch (needs FLUX.1 Kontext dev), Restore and enlarge photos (needs FLUX.1 Kontext dev), Character sheet, 12 poses (needs FLUX.1 dev), 6-panel comic page (needs FLUX.1 dev), Long-form article batch (needs Llama 3.3 70B).

Rent or buy?

This card is $1,500 to buy. The cheapest listed rate on Vast.ai is $0.230/hour, but that is the floor: we budget $0.276/hour, a 20% premium, because idle time, storage and unavailable cheap instances all land on the same bill. At that rate buying wins after 5,435 GPU-hours. Below it you are paying for idle silicon.

How you would use it	GPU-hours a year	Rental cost a year	Time to break even
2 hours a day, hobby	730	$201	7.4 years
8 hours a day, working on it	2,920	$806	1.9 years
24/7, always-on agent	8,760	$2,418	7.4 months

At hobby usage this card is very unlikely to pay for itself before it is superseded. Rent it. Rental figures include a 20% premium over the cheapest listed rate. Ignores electricity, resale and the fact that a rented card can be a newer one tomorrow.

Rental price

$0.230/hr+12.7% since 2026-10-01low $0.204 · high $0.230

Cheapest of the RunPod and Vast on-demand rates we see, sampled daily. Spot and interruptible pricing runs lower.