Playground v2.5 · 23 GPUs measured first-party · text-to-image · Updated October 2026
What GPU Do You Need for Playground v2.5?
We ran Playground v2.5 at 1024px on 23 GPUs, from budget cards to datacenter parts, and measured image speed, power draw and peak VRAM on every one. The fastest was NVIDIA H100 80GB HBM3 at 21.48 images/min.
Benchmarked weights: playgroundai/playground-v2.5-1024px-aesthetic
NVIDIA H100 80GB HBM3
21.48 images/min on Playground v2.5, the ceiling. Measured on our bench. 80GB of VRAM, $30,000 at launch.
- 21.48 images/min on Playground v2.5
- 80GB, clears the Playground v2.5 floor
- Rentable by the hour rather than bought
- 700W board rating
- Datacenter or workstation hardware, not a retail purchase
NVIDIA GeForce RTX 5090
13.28 images/min on Playground v2.5, fastest card you can buy at retail. Measured on our bench. 32GB of VRAM, $1,999 at launch.
- 13.28 images/min on Playground v2.5
- 32GB, clears the Playground v2.5 floor
- 575W board rating
NVIDIA GeForce RTX 3060
1.76 images/min on Playground v2.5, lowest launch price that still fits. Measured on our bench. 12GB of VRAM, $329 at launch.
- 1.76 images/min on Playground v2.5
- 12GB, clears the Playground v2.5 floor
- 170W board rating
NVIDIA GeForce RTX 4070 Super
4.79 images/min on Playground v2.5, most speed per dollar. Measured on our bench. 12GB of VRAM, $599 at launch. That is 8.0 images/min per $1,000 of launch price.
- 4.79 images/min on Playground v2.5
- 12GB, clears the Playground v2.5 floor
- 220W board rating
What GPU Do You Need for Playground v2.5?, images/min by GPU
Top 15 shown; 8 more cards in the full table below.
Efficiency: images/min per 100W drawn
Top 15 shown; 8 more cards in the full table below.
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
Top 15 shown; 8 more cards in the full table below.
Launch price, not street price, so it ages. A speed leaderboard always crowns the most expensive card; this is the counterweight.
Playground v2.5. Measured image speed by GPU
| GPU | Images/min | s per image | img/W·min | Avg power |
|---|---|---|---|---|
| NVIDIA H100 80GB HBM3 | 21.48 | 2.79 | 0.04 | 542.8 W |
| NVIDIA GeForce RTX 5090 | 13.28 | 4.518 | 0.025 | 523.0 W |
| NVIDIA L40S | 10.85 | 5.53 | 0.033 | 328.4 W |
| NVIDIA GeForce RTX 4090 | 9.96 | 6.021 | 0.025 | 396.1 W |
| GeForce RTX 5080 | 7.6 | 7.897 | 0.025 | 309.4 W |
| GeForce RTX 4080 Super | 6.74 | 8.897 | 0.029 | 233.2 W |
| NVIDIA GeForce RTX 4080 | 6.33 | 9.48 | 0.024 | 262.1 W |
| GeForce RTX 5070 Ti | 5.88 | 10.206 | 0.024 | 247.4 W |
| NVIDIA GeForce RTX 4070 Ti Super | 5.57 | 10.769 | 0.025 | 220.8 W |
| NVIDIA RTX PRO 4000 Blackwell | 5.54 | 10.833 | 0.038 | 144.8 W |
| NVIDIA GeForce RTX 3090 Ti | 5.39 | 11.123 | 0.014 | 394.2 W |
| NVIDIA GeForce RTX 4070 Ti | 5.29 | 11.343 | 0.022 | 238.9 W |
| NVIDIA GeForce RTX 3080 Ti | 4.9 | 12.243 | 0.014 | 346.5 W |
| NVIDIA GeForce RTX 3090 | 4.79 | 12.523 | 0.014 | 346.4 W |
| NVIDIA GeForce RTX 4070 Super | 4.79 | 12.537 | 0.023 | 206.0 W |
| GeForce RTX 5070 | 4.37 | 13.724 | 0.024 | 179.8 W |
| NVIDIA RTX 4000 (Ada Generation) | 4.32 | 13.875 | 0.035 | 123.5 W |
| NVIDIA GeForce RTX 4070 | 4.01 | 14.951 | 0.02 | 197.4 W |
| NVIDIA RTX A4000 | 3.21 | 18.672 | 0.023 | 139.6 W |
| NVIDIA L4 | 3.01 | 19.97 | 0.042 | 72.0 W |
| GeForce RTX 5060 Ti | 2.93 | 20.499 | 0.032 | 90.3 W |
| NVIDIA GeForce RTX 4060 Ti 16GB | 2.89 | 20.727 | 0.018 | 157.8 W |
| NVIDIA GeForce RTX 3060 | 1.76 | 34.077 | 0.012 | 151.5 W |
What the numbers show. Across 23 GPUs measured on our own bench, H100 80GB HBM3 is fastest at 21.5 images/min. L4 is the most efficient, 3.01 images/min at 72W (0.042 images/min per watt). Per dollar of launch price, RTX 5060 Ti gives the most (8.9 images/min per $1,000). The fastest card with 16GB or less is RTX 5080 at 7.60 images/min. The measured peak was ~11GB, so it runs on 12GB cards and up. That floor is a property of the model, so it applies to every GPU, measured or not.
About Playground v2.5. Playground v2.5: from playgroundai, 2.6B parameters, on Hugging Face since February 2024. 285,945 downloads in the last 30 days.
How it compares. L40S: Playground v2.5 10.85 images/min, SDXL Turbo 423.5 (3B), Stable Diffusion 3.5 Medium 11.58 (2B), FLUX.2 klein 4B 50.29 (4B), Stable Diffusion 1.5 58.54. All 4 beat Playground v2.5 here.
Cost on a rented GPU. 1,000 images of Playground v2.5: $0.34 on a RTX 3060 ($0.036/hr, 9.5 hours), $1.66 on a H100 80GB HBM3 ($2.14/hr, 47 min, 4.9x the cost).
Playground v2.5: 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 | 1.76 | $0.34 |
| NVIDIA RTX A4000 | $0.078/hr | 3.21 | $0.40 |
| GeForce RTX 5070 Ti | $0.15/hr | 5.88 | $0.42 |
| NVIDIA GeForce RTX 3090 | $0.12/hr | 4.79 | $0.42 |
| GeForce RTX 5080 | $0.21/hr | 7.6 | $0.46 |
| NVIDIA GeForce RTX 4070 Ti Super | $0.16/hr | 5.57 | $0.48 |
| NVIDIA GeForce RTX 5090 | $0.39/hr | 13.28 | $0.49 |
| NVIDIA GeForce RTX 4070 | $0.12/hr | 4.01 | $0.51 |
| NVIDIA GeForce RTX 4080 | $0.20/hr | 6.33 | $0.53 |
| GeForce RTX 4080 Super | $0.22/hr | 6.74 | $0.55 |
| NVIDIA GeForce RTX 4070 Ti | $0.18/hr | 5.29 | $0.55 |
| NVIDIA GeForce RTX 4090 | $0.34/hr | 9.96 | $0.56 |
| NVIDIA GeForce RTX 3080 Ti | $0.18/hr | 4.9 | $0.61 |
| NVIDIA RTX PRO 4000 Blackwell | $0.20/hr | 5.54 | $0.61 |
| GeForce RTX 5070 | $0.19/hr | 4.37 | $0.72 |
| NVIDIA RTX 4000 (Ada Generation) | $0.20/hr | 4.32 | $0.77 |
| GeForce RTX 5060 Ti | $0.14/hr | 2.93 | $0.77 |
| NVIDIA GeForce RTX 3090 Ti | $0.27/hr | 5.39 | $0.83 |
| NVIDIA L40S | $0.79/hr | 10.85 | $1.21 |
| NVIDIA H100 80GB HBM3 | $2.14/hr | 21.48 | $1.66 |
| NVIDIA L4 | $0.44/hr | 3.01 | $2.44 |
Cheapest hourly rate we track on RunPod and Vast.ai, divided by the measured speed. Startup time and storage are extra.
Speed tiers for Playground v2.5. 6-30 images/min: 7 (RTX 5090, RTX 4090, RTX 5080); under 6 images/min: 16 (RTX 5070 Ti, RTX 4070 Ti Super, RTX 3090 Ti). 30 images/min means two seconds or less per picture.
Time per image. Playground v2.5 at 1024px and 50 steps: 2.8s per image on the H100 80GB HBM3, 34.1s on the RTX 3060. A batch of 100 takes 5 min on the fastest card and 57 min on the slowest.
VRAM for Playground v2.5. Measured peak 9.9GB, so 12GB is the smallest common card size; smallest card it ran on: RTX 4070 Ti (12GB).
Power on Playground v2.5. Most efficient: L4, 72W, 0.40 kWh per 1,000 images. Hungriest: H100 80GB HBM3, 543W, 0.42 kWh. At $0.15/kWh: $0.060 per 1,000 images.
Our verdict
Fastest on Playground v2.5: NVIDIA H100 80GB HBM3, 21.48 images/min. Best desktop card: NVIDIA GeForce RTX 5090, 13.28 images/min. Cheapest consumer card that ran it: NVIDIA GeForce RTX 3060 ($329, 1.76 images/min). Cheapest to rent per job: NVIDIA GeForce RTX 3060, $0.34 per 1,000 images.
FAQ
- What GPU do I need to run Playground v2.5?
- About 10GB. Cheapest consumer card that ran it: NVIDIA GeForce RTX 3060 (12GB, 1.76 images/min).
- How fast is Playground v2.5 on the NVIDIA GeForce RTX 5090?
- 13.28 images/min at 523W, 62% of the NVIDIA H100 80GB HBM3.
- How much does it cost to run Playground v2.5 in the cloud?
- $0.34 per 1,000 images on a NVIDIA GeForce RTX 3060 at $0.036/hr, cheapest of 21 rentable cards we measured.
- Can I run Playground v2.5 on a 12GB, 16GB or 24GB card?
- It used 9.9GB at the precision we tested. 12GB: yes; 16GB: yes; 24GB: yes.
- Is the RTX 4090 or the RTX 3090 faster for Playground v2.5?
- The RTX 4090: 9.96 vs 4.79 images/min, 108% faster on our bench.
- Should I buy or rent a RTX 5090 for Playground v2.5?
- Its $1,999 launch price buys 5,139 rented hours at $0.39/hr, enough for about 4,095x 1,000 images of Playground v2.5. Buy only if you'll run more than that.
How we test
Playground v2.5 at 1024px 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.