NVIDIA GeForce RTX 5090, AI & Machine Learning Benchmarks & Specs

32GB · AI Score 22.3/100 · first-party measured on 12 AI workloads

22.3 AI Score ✓ Measured

Every number on this page is first-party: NVIDIA GeForce RTX 5090 was run on our pinned 12-workload AI suite on 2026-07-11, with under 0.5% run-to-run variance. On Llama 3.1 8B (Q4_K_M) NVIDIA GeForce RTX 5090 delivers about 268.14 tokens/sec. Stepping up to Qwen3 32B it holds roughly 71.15 tok/s. Llama 3.3 70B does not fit. It needs roughly 42GB and this card has 32GB. For image generation, SDXL runs at 10.54 it/s, and FLUX.1-dev at 0.95 it/s. 2 of the 12 workloads won't fit on 32GB at the tested precision, Llama 3.3 70B, Qwen-Image-Edit. We publish those as hard gates rather than quietly dropping to a smaller quant. NVIDIA GeForce RTX 5090 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 37

Llama 3.2 1B1060.14

Qwen3 0.6B868.61

LFM2.5 2.6B566.57

MiniCPM5 2B507.31

gpt-oss-20b415.12

Nemotron 3 Nano 4B410.94

Nemotron 3.5 Lightning 30B A3B376.06

Qwen3 4B375.6

Granite 4.1 3B375.37

Qwen3-Coder 30B A3B365.84

Qwen3 30B A3B Instruct 2507352.33

Spark-X2.5-4B352.29

Workload	Result	Telemetry	Data
Qwen3 0.6B	868.61 tok/s	118 W51°CQ4_K_M	✓ Measured
Llama 3.2 1B	1060.14 tok/s	119 W51°CQ4_K_M	✓ Measured
MiniCPM5 2B	507.31 tok/s	175 W51°CQ4_K_M	✓ Measured
LFM2.5 2.6B	566.57 tok/s	142 W53°CQ4_K_M	✓ Measured
Granite 4.1 3B	375.37 tok/s	190 W58°CQ4_K_M	✓ Measured
Agents-A1-4B	328.96 tok/s	245 W54°CQ4_K_M	✓ Measured
Nemotron 3 Nano 4B	410.94 tok/s	219 W54°CQ4_K_M	✓ Measured
Qwen3 4B	375.6 tok/s	3 GB peak87 W48°C4.32 tok/WQ4_K_M	✓ Measured
Spark-X2.5-4B	352.29 tok/s	264 W58°CQ4_K_M	✓ Measured
DeepSeek Coder 7B Instruct v1.5	291.38 tok/s	266 W62°CQ4_K_M	✓ Measured
OLMo 3 7B Instruct	266.68 tok/s	314 W56°CQ4_K_M	✓ Measured
OLMo 3 7B Think	266.72 tok/s	288 W55°CQ4_K_M	✓ Measured
Qwen2-7B-Instruct	277.64 tok/s	290 W56°CQ4_K_M	✓ Measured
Qwen2.5-7B	284.6 tok/s	301 W59°CQ4_K_M	✓ Measured
Qwen2.5-Coder 7B	284.59 tok/s	265 W60°CQ4_K_M	✓ Measured
Apertus-8B-Instruct	257.69 tok/s	319 W55°CQ4_K_M	✓ Measured
Llama 3 8B	259.18 tok/s	297 W56°CQ4_K_M	✓ Measured
Llama 3.1 8B	268.14 tok/s	5 GB peak94 W52°C2.85 tok/WQ4_K_M	✓ Measured
Qwen3 8B	243.91 tok/s	265 W45°CQ4_K_M	✓ Measured
Nemotron Nano 9B v2	202.68 tok/s	291 W58°CQ4_K_M	✓ Measured
Ornith 1.5 9B	225.79 tok/s	305 W56°CQ4_K_M	✓ Measured
Gemma 4 12B	151.54 tok/s	278 W45°CQ4_K_M	✓ Measured
Qwen2.5-Coder 14B	149.75 tok/s	8.6 GB peak117 W54°C1.28 tok/WQ4_K_M	✓ Measured
Qwen3 14B	143.11 tok/s	314 W46°CQ4_K_M	✓ Measured
gpt-oss-20b	415.12 tok/s	203 W53°CQ4_K_M	✓ Measured
Gemma 4 26B A4B	264.69 tok/s	215 W52°CQ4_K_M	✓ Measured
Qwen3.6 27B	76.6 tok/s	426 W57°CQ4_K_M	✓ Measured
Qwen3.8 27B	75.74 tok/s	413 W57°CQ4_K_M	✓ Measured
Nemotron 3.5 Lightning 30B A3B	376.06 tok/s	178 W55°CQ4_K_M	✓ Measured
Qwen3 30B A3B	344.67 tok/s	156 W39°CQ4_K_M	✓ Measured
Qwen3 30B A3B Instruct 2507	352.33 tok/s	189 W56°CQ4_K_M	✓ Measured
Qwen3-Coder 30B A3B	365.84 tok/s	197 W54°CQ4_K_M	✓ Measured
Gemma 4 31B	71.42 tok/s	439 W59°CQ4_K_M	✓ Measured
Qwen3 32B	71.15 tok/s	19 GB peak83 W54°C0.86 tok/WQ4_K_M	✓ Measured
Ornith 1.5 35B A3B	296.44 tok/s	188 W52°CQ4_K_M	✓ Measured
Qwen3.6 35B A3B	300.13 tok/s	210 W54°CQ4_K_M	✓ Measured
Llama 3.3 70B	✕ Won't fit needs ~46 GB	VRAM-gated at this precision	✓ Measured

Image Generation images/min 17

SDXL Turbo902.04

DreamShaper XL Lightning122.85

Stable Diffusion 1.5102.36

DreamShaper XL Turbo70.72

Stable Diffusion 2.165.53

FLUX.2 klein 4B59.33

Sana 1.6B51.52

PixArt-Sigma XL30.01

Stable Diffusion XL21.08

SSD-1B20.49

Kolors13.53

Playground v2.513.28

Workload	Result	Telemetry	Data
Stable Diffusion 1.5	102.36 images/min	451 W57°C	✓ Measured
Stable Diffusion 2.1	65.53 images/min	309 W47°C	✓ Measured
SDXL Turbo	902.04 images/min	33 W51°C	✓ Measured
SSD-1B	20.49 images/min	589 W66°C	✓ Measured
Sana 1.6B	51.52 images/min	549 W64°C	✓ Measured
Stable Diffusion XL	21.08 images/min	14.7 GB peak525 W57°C2.9 s/img	✓ Measured
DreamShaper XL Lightning	122.85 images/min	415 W62°C	✓ Measured
DreamShaper XL Turbo	70.72 images/min	535 W50°C	✓ Measured
Playground v2.5	13.28 images/min	523 W63°C	✓ Measured
PixArt-Sigma XL	30.01 images/min	547 W64°C	✓ Measured
FLUX.2 klein 4B	59.33 images/min	518 W60°C	✓ Measured
Kolors	13.53 images/min	593 W65°C	✓ Measured
Z-Image	1.72 images/min	525 W75°C	✓ Measured
AuraFlow v0.3	4.35 images/min	572 W74°C	✓ Measured
Z-Image Turbo	10.73 images/min	25.9 GB peak563 W64°C5.6 s/img	✓ Measured
Chroma1-HD	1.29 images/min	521 W68°C	✓ Measured
FLUX.1 dev	2.04 images/min	30.7 GB peak288 W68°C29.4 s/img	✓ Measured CPU offload

Image Editing images/min 1

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

Image to Video clips/min 5

LTX-Video (image to video)6.152

Stable Video Diffusion2.836

Wan 2.2 TI2V-5B (image to video)2.284

Stable Video Diffusion XT1.344

CogVideoX-5B I2V0.447

Workload	Result	Telemetry	Data
Stable Video Diffusion	2.84 clips/min	562 W79°C	✓ Measured
LTX-Video (image to video)	6.15 clips/min	530 W75°C	✓ Measured
Wan 2.2 TI2V-5B (image to video)	2.28 clips/min	564 W70°C	✓ Measured 2 hosts ±0%
Stable Video Diffusion XT	1.34 clips/min		✓ Measured 2 hosts ±5% · CPU offload
CogVideoX-5B I2V	0.45 clips/min		✓ Measured 2 hosts ±1% · CPU offload

Video Generation frames/s 5

LTX-Video (distilled)10.449

Wan 2.1 1.3B0.879

CogVideoX-2B0.7

Wan 2.2 5B (720p)0.588

CogVideoX-5B0.259

Workload	Result	Telemetry	Data
Wan 2.1 1.3B	0.88 frames/s	566 W75°C52.5 s/clip	✓ Measured 2 hosts ±6%
CogVideoX-2B	0.7 frames/s	70 s/clip	✓ Measured 2 hosts ±3%
CogVideoX-5B	0.26 frames/s	572 W78°C189 s/clip	✓ Measured
LTX-Video (distilled)	10.45 frames/s	448 W50°C9.3 s/clip	✓ Measured
Wan 2.2 5B (720p)	0.59 frames/s	456 W69°C83.4 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 2026-07-11 · harness 2.0.0-standalone.

NVIDIA GeForce RTX 5090 specifications

Architecture	Blackwell (GB202)
CUDA cores	21,760
VRAM	32GB GDDR7
Memory bus	512-bit
Memory bandwidth	1792 GB/s
Boost clock	2,407 MHz
TDP	575 W
Process	5nm
Interface	PCIe 5.0 x16
Release date	2025-01-30
Launch MSRP	$1,999

Verdict, capable, but 32GB sets the ceiling

NVIDIA GeForce RTX 5090 scores 22.3/100, #19 of 102. It ran 7 of 12; 2 exceeded its 32GB. Every figure here is our own measurement.

Relative performance: where the NVIDIA GeForce RTX 5090 lands

100% = this card, AI & Machine Learning headline metric (AI Score). #2 of 61 desktop cards in this vertical.

GPU	%	AI Score
NVIDIA RTX 6000 Ada Generation	107%	23.8
NVIDIA GeForce RTX 5090	100%	22.3
NVIDIA RTX 5880 Ada Generation	66%	14.7
NVIDIA RTX 5000 Ada Generation	49%	11
NVIDIA GeForce RTX 4090	48%	10.6
NVIDIA GeForce RTX 3090 Ti	38%	8.5

One tier up: NVIDIA RTX 6000 Ada Generation (+7%) · compare head-to-head
One tier down: NVIDIA RTX 5880 Ada Generation (-34%) · compare head-to-head

Same card, other workloads: NVIDIA GeForce RTX 5090 Gaming benchmarks

← All AI & Machine Learning GPU rankings

The silicon

Transistors	92,200 million
Die size	750 mm²
Process node	4 nm
Fabricated by	TSMC
Transistor density	122.9 million per mm²

Denser than 96% 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
Full codebase review	6.7 min	13.03 Wh	measured
Animate a batch of images	8.8 min	82.24 Wh	measured
60-second AI short film	9.3 min	23.85 Wh	all 3 stages measured
24-frame storyboard	12.5 min	21.46 Wh	all 2 stages measured
Short social clips	24.8 min	114.49 Wh	all 3 stages measured

Can't run: Long-form article batch (needs Llama 3.3 70B).

Rent or buy?

This card is $1,999 to buy. The cheapest listed rate on Vast.ai is $0.463/hour, but that is the floor: we budget $0.556/hour, a 20% premium, because idle time, storage and unavailable cheap instances all land on the same bill. At that rate buying wins after 3,598 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	$406	4.9 years
8 hours a day, working on it	2,920	$1,622	1.2 years
24/7, always-on agent	8,760	$4,867	4.9 months

At steady usage this card pays for itself inside a normal ownership window. 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

$0.463/hr+72.1% since 2026-08-14low $0.213 · high $0.482

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