SD Turbo · 7 GPUs measured first-party · text-to-image · Updated October 2026
SD Turbo on 7 GPUs, measured first-party: NVIDIA GeForce RTX 4090 leads at 786.1 images/min, RTX 3060 trails at 268 images/min, and it peaked at 4GB of VRAM.
Benchmarked weights: stabilityai/sd-turbo

786.1 images/min on SD Turbo, the ceiling. Measured on our bench. 24GB of VRAM, $1,599 at launch.

644.3 images/min on SD Turbo, fastest card you can buy at retail. Measured on our bench. 16GB of VRAM, $749 at launch.

267.9 images/min on SD Turbo, lowest launch price that still fits. Measured on our bench. 12GB of VRAM, $329 at launch.

535.1 images/min on SD Turbo, most speed per dollar. Measured on our bench. 12GB of VRAM, $549 at launch. That is 974.6 images/min per $1,000 of launch price.
What GPU Do You Need for SD Turbo?, images/min by GPU
Efficiency: images/min per 100W drawn
Power is the average pulled during the run, sampled at 1Hz. The fastest card is often not the one here, and for anything left running this is the number that shows up on the bill.
Value: images/min per $1,000 of MSRP
Launch price, not street price, so it ages. A speed leaderboard always crowns the most expensive card; this is the counterweight.
SD Turbo. Measured image generation speed by GPU
| GPU | Images/min | s per image | img/W·min | Avg power |
|---|---|---|---|---|
| NVIDIA GeForce RTX 4090 | 786.1 | 0.076 | 8.668 | 90.7 W |
| GeForce RTX 5070 Ti | 644.3 | 0.093 | 10.359 | 62.2 W |
| GeForce RTX 5070 | 535.1 | 0.112 | 8.644 | 61.9 W |
| NVIDIA GeForce RTX 4060 Ti 16GB | 361.3 | 0.166 | 6.935 | 52.1 W |
| NVIDIA GeForce RTX 3070 Founders Edition | 357.8 | 0.168 | 8.132 | 44.0 W |
| NVIDIA GeForce RTX 3060 Ti | 327.8 | 0.183 | 3.998 | 82.0 W |
| NVIDIA GeForce RTX 3060 | 267.9 | 0.224 | 4.392 | 61.0 W |
What the numbers show. Across 7 GPUs measured on our own bench, RTX 4090 is fastest at 786 images/min. The slowest, RTX 3060, manages 268, so the spread is 2.9x from top to bottom. RTX 5070 Ti is the most efficient, 644 images/min at 62W. Per dollar of launch price, RTX 5070 gives the most (974.6 images/min per $1,000). The fastest card with 16GB or less is RTX 5070 Ti at 644 images/min.
About SD Turbo. SD Turbo: from stabilityai, 0.9B parameters, on Hugging Face since November 2023. 613,825 downloads in the last 30 days.
How it compares. RTX 4090: SD Turbo 786.1 images/min, Stable Diffusion 2.1 60.58, Stable Diffusion 1.5 80.14, LCM DreamShaper v7 260.9, PixArt-Sigma XL 24.27. SD Turbo beats all 4 here.
Cost on a rented GPU. 1,000 images of SD Turbo: $0.002 on a RTX 3060 ($0.036/hr, 4 min), $0.007 on a RTX 4090 ($0.34/hr, 1 min, 3.2x the cost).
SD Turbo: cost per 1,000 images on rented GPUs
| GPU | Cheapest rate | Speed (images/min) | Cost per 1,000 images |
|---|---|---|---|
| NVIDIA GeForce RTX 3060 | $0.036/hr | 267.9 | $0.002 |
| NVIDIA GeForce RTX 3070 Founders Edition | $0.075/hr | 357.8 | $0.003 |
| GeForce RTX 5070 Ti | $0.15/hr | 644.3 | $0.004 |
| GeForce RTX 5070 | $0.19/hr | 535.1 | $0.006 |
| NVIDIA GeForce RTX 4090 | $0.34/hr | 786.1 | $0.007 |
Cheapest hourly rate we track on RunPod and Vast.ai, divided by the measured speed. Startup time and storage are extra.
Speed tiers for SD Turbo. 30+ images/min: 7 (RTX 4090, RTX 5070 Ti, RTX 5070). 30 images/min means two seconds or less per picture.
Time per image. SD Turbo at 512px and 1 steps: 0.1s per image on the RTX 4090, 0.2s on the RTX 3060. A batch of 100 takes 0 min on the fastest card and 0 min on the slowest.
VRAM for SD Turbo. Measured peak 3.8GB, so 8GB is the smallest common card size; smallest card it ran on: RTX 3070 Founders Edition (8GB).
Power on SD Turbo. Most efficient: RTX 5070 Ti, 62W, 1.6 Wh per 1,000 images. Hungriest: RTX 4090, 91W, 1.9 Wh.
Fastest on SD Turbo: NVIDIA GeForce RTX 4090, 786.1 images/min. Cheapest consumer card that ran it: NVIDIA GeForce RTX 3060 ($329, 267.9 images/min). Cheapest to rent per job: NVIDIA GeForce RTX 3060, $0.002 per 1,000 images.
SD Turbo at 512px in diffusers, bf16, native precision with no offload, timed over three generations after a warmup, with power and VRAM sampled throughout. Image speed tracks tensor compute and architecture generation more than memory bandwidth, so the order here differs from our LLM boards.