AI image upscaling · 1 model · 11 GPUs measured first-party · Updated October 2026
Best GPU for AI image upscaling
Which graphics card to use for AI image upscaling, from first-party measurements of Swin2SR 4x Upscaler on 11 GPUs.
NVIDIA B300
58.29 images/min on Swin2SR 4x Upscaler, the ceiling. Measured on our bench. 288GB of VRAM, $40,000 at launch.
- 58.29 images/min on Swin2SR 4x Upscaler
- 288GB, clears the Swin2SR 4x Upscaler floor
- Rentable by the hour rather than bought
- 1400W board rating
- Datacenter or workstation hardware, not a retail purchase
NVIDIA RTX PRO 6000 Blackwell Workstation Edition
52.83 images/min on Swin2SR 4x Upscaler, lowest launch price that still fits. Measured on our bench. 96GB of VRAM, $8,565 at launch.
- 52.83 images/min on Swin2SR 4x Upscaler
- 96GB, clears the Swin2SR 4x Upscaler floor
- Rentable by the hour rather than bought
- 600W board rating
- Datacenter or workstation hardware, not a retail purchase
AI upscalers enlarge an image (4x here) and invent plausible detail instead of blurring it. They are the last step of most image pipelines and a common batch job for photo and game-texture work.
We measured 1 model for AI image upscaling on 11 GPUs. Speed is images upscaled per minute. Every number below is a first-party run on our own harness; cards absent from a model's chart have not been run on it yet.
Swin2SR 4x Upscaler: images/min by GPU
Which models fit which card, for AI image upscaling
| Model | VRAM used | 8GB card | 12GB card | 16GB card | 24GB card | 32GB card | Licence |
|---|---|---|---|---|---|---|---|
| Swin2SR 4x Upscaler | 6.3GB | Yes | Yes | Yes | Yes | Yes | Apache-2.0 |
From the lowest VRAM peak we measured for each model, plus 5% headroom. 'No' means it did not fit in that much memory at our settings, not that no setting ever could.
Swin2SR 4x Upscaler: every GPU we measured
| GPU | images/min | VRAM | Power |
|---|---|---|---|
| NVIDIA B300 | 58.29 | 288GB | 437.0 W |
| NVIDIA RTX PRO 6000 Blackwell Workstation Edition | 52.83 | 96GB | 369.8 W |
| NVIDIA L40S | 34.14 | 48GB | 247.1 W |
| NVIDIA A100 80GB SXM4 | 28.72 | 80GB | 184.0 W |
| NVIDIA H200 | 25.85 | 141GB | 214.3 W |
| NVIDIA H100 80GB HBM3 | 25.48 | 80GB | 203.0 W |
| NVIDIA A10G | 22.31 | 24GB | 147.6 W |
| NVIDIA L4 | 18.23 | 24GB | 71.7 W |
| NVIDIA A100 40GB SXM4 | 17.25 | 40GB | 116.4 W |
| NVIDIA B200 | 11.97 | 192GB | 281.6 W |
| NVIDIA T4 | 10.53 | 16GB | 65.4 W |
What the numbers show.
Swin2SR 4x Upscaler: fastest on the NVIDIA B300 at 58.29 images/min, 5.54x the slowest card we measured (NVIDIA T4); it used about 6.3GB of VRAM.
How it compares. L40S: Swin2SR 4x Upscaler 34.14 images/min, PixArt-Sigma XL 26.95, FLUX.1 Schnell 26.24 (12B), FLUX.2 klein 9B 25.95 (9B), Z-Image Turbo (1024px) 18.2 (6B). Swin2SR 4x Upscaler beats all 4 here.
Cost on a rented GPU. 1,000 images of Swin2SR 4x Upscaler: $0.22 on a T4 ($0.14/hr, 95 min), $1.98 on a B300 ($6.94/hr, 17 min, 9.2x the cost).
Swin2SR 4x Upscaler: cost per 1,000 images on rented GPUs
| GPU | Cheapest rate | Speed (images/min) | Cost per 1,000 images |
|---|---|---|---|
| NVIDIA T4 | $0.14/hr | 10.53 | $0.22 |
| NVIDIA RTX PRO 6000 Blackwell Workstation Edition | $1.08/hr | 52.83 | $0.34 |
| NVIDIA L40S | $0.79/hr | 34.14 | $0.39 |
| NVIDIA L4 | $0.44/hr | 18.23 | $0.40 |
| NVIDIA A100 40GB SXM4 | $0.47/hr | 17.25 | $0.46 |
| NVIDIA A100 80GB SXM4 | $0.95/hr | 28.72 | $0.55 |
| NVIDIA H100 80GB HBM3 | $2.14/hr | 25.48 | $1.40 |
| NVIDIA B300 | $6.94/hr | 58.29 | $1.98 |
| NVIDIA H200 | $3.59/hr | 25.85 | $2.31 |
| NVIDIA B200 | $5.98/hr | 11.97 | $8.33 |
Cheapest hourly rate we track on RunPod and Vast.ai, divided by the measured speed. Startup time and storage are extra.
Speed tiers for Swin2SR 4x Upscaler. 30+ images/min: 3 (B300, RTX PRO 6000 Blackwell Workstation Edition, L40S); 6-30 images/min: 8 (A100 80GB SXM4, H200, H100 80GB HBM3). 30 images/min means two seconds or less per picture.
Time per image. Swin2SR 4x Upscaler: 1.0s per image on the B300, 5.7s on the T4. A batch of 100 takes 2 min on the fastest card and 9 min on the slowest.
VRAM for Swin2SR 4x Upscaler. Measured peak 6.0GB, so 8GB is the smallest common card size; smallest card it ran on: T4 (16GB).
Power on Swin2SR 4x Upscaler. Most efficient: L4, 72W, 65.6 Wh per 1,000 images. Hungriest: B300, 437W, 0.12 kWh.
Our verdict
For AI image upscaling, the NVIDIA B300 is the fastest card we measured. We have not measured a consumer card on this job yet; the picks above are datacenter and workstation hardware. Check the fit table before buying: VRAM, not speed, is what rules a card out.
FAQ
- What is the fastest GPU for AI image upscaling?
- In our runs, the NVIDIA B300 at 58.29 images/min on Swin2SR 4x Upscaler. We measured 1 model on 11 GPUs for this page.
- How much VRAM do I need for AI image upscaling?
- The lightest model here, Swin2SR 4x Upscaler, used about 6.3GB. The table above shows which models fit 8, 12, 16, 24 and 32GB cards, from measured peaks.
- Are these numbers measured or estimated?
- Measured. Every number on this page is a first-party run on our own harness, with power and VRAM sampled during the run. Cards we have not run yet are simply absent, not filled in.
- What GPU do I need to run Swin2SR 4x Upscaler?
- About 6GB. Smallest card that ran it: NVIDIA T4 (16GB).
- How much does it cost to run Swin2SR 4x Upscaler in the cloud?
- $0.22 per 1,000 images on a NVIDIA T4 at $0.14/hr, cheapest of 10 rentable cards we measured.
- Can I run Swin2SR 4x Upscaler on a 12GB, 16GB or 24GB card?
- It used 6.0GB at the precision we tested. 12GB: yes; 16GB: yes; 24GB: yes.
- Is the A100 80GB SXM4 or the H100 80GB HBM3 faster for Swin2SR 4x Upscaler?
- The A100 80GB SXM4: 28.72 vs 25.48 images/min, 13% faster on our bench.
How we test
Each model runs a fixed workload on every card: a warmup, then timed runs with power, temperature and VRAM sampled every half second through NVML. Models run at the precision and settings from their model card. Datacenter cards run on Modal; consumer cards on rented machines. Non-commercially licensed models are not part of this page.