// tools/speed-benchmark-table

Diffusion Model Speed Benchmark Table

Images per second and cost efficiency for Flux Schnell, Flux Dev, and SDXL across 11 GPU tiers. Toggle between speed and efficiency views to find the right GPU for your workload. 11 GPUs · 3 models · benchmarks June 2026

View mode

Filter by tier

Key insights

For real-time UX (latency matters): H100 or H200 — fastest per request. Use the imgs/sec view above.
For batch processing (throughput matters): check the imgs/$ column — cheapest GPUs often win on cost efficiency. The RTX 3090 at $0.09/hr delivers 22K Flux Schnell images per dollar.
Runflow uses a warm GPU fleet (RTX 4090 to H200). For throughput at scale, they handle GPU selection automatically.

GPU↕	Tier	$/hr↕	Flux Schnell	Flux Dev	SDXL
			imgs/sec↓	imgs/sec↕	imgs/sec↕
H200 SXM 141GB VRAM · Vast.ai	marketplace	$3.88	3.89	0.64	0.78
H100 NVL 94GB VRAM · Vast.ai	marketplace	$2.52	3.06	0.50	0.61
H100 80GBbest efficiency 80GB VRAM · Salad	community	$0.99	2.78	0.44	0.56
RTX 5090 32GB VRAM · Vast.ai	marketplace	$0.87	1.94	0.31	0.39
A100 80GB 80GB VRAM · TensorDock	marketplace	$0.75	1.67	0.26	0.33
L40S 48GB VRAM · Vast.ai	marketplace	$0.47	1.06	0.17	0.21
RTX 4090best all-around 24GB VRAM · Salad	community	$0.20	0.97	0.15	0.19
RTX 4080 16GB VRAM · Vast.ai	marketplace	$0.33	0.72	0.11	0.14
RTX 3090 24GB VRAM · Salad	community	$0.09	0.56	0.09	0.11
A40 48GB VRAM · Vast.ai	marketplace	$0.29	0.56	0.09	0.12
RTX A6000 48GB VRAM · Vast.ai	marketplace	$0.39	0.56	0.09	0.12

Color coding:Top 3 in selected metricBottom 3 in selected metricBest all-around (RTX 4090)H100 (fastest raw speed)

Flux Schnell (4-step) top 3

1H200 SXM

3.89/s

2H100 NVL

3.06/s

3H100 80GB

2.78/s

Flux Dev (28-step) top 3

1H200 SXM

0.64/s

2H100 NVL

0.50/s

3H100 80GB

0.44/s

SDXL (30-step) top 3

1H200 SXM

0.78/s

2H100 NVL

0.61/s

3H100 80GB

0.56/s

Methodology

Throughput estimated from single-GPU benchmarks, batch mode, 1024×1024. Community and marketplace tiers use median pricing from the cheapest listing available as of June 2026. imgs/sec = imgsPerHr ÷ 3600. imgs/$ = imgsPerHr ÷ price per hour. Real-world numbers may vary based on model weights, scheduler settings, and host network latency.

Throughput estimated from single-GPU benchmarks, batch mode, 1024×1024 · Community and marketplace tiers use median pricing · GPU Cost Calculator → · GPU pricing breakdowns → · Updated June 2026