Image segmentation with SAM · 2 models · 11 GPUs measured first-party · Updated October 2026
Best GPU for Image segmentation with SAM
Which graphics card to use for image segmentation with SAM, from first-party measurements of SAM ViT-Base, SAM ViT-Huge on 11 GPUs.
NVIDIA B300
1719.6 images/min on SAM ViT-Base, the ceiling. Measured on our bench. 288GB of VRAM, $40,000 at launch.
- 1719.6 images/min on SAM ViT-Base
- 288GB, clears the SAM ViT-Base 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
1079.9 images/min on SAM ViT-Base, lowest launch price that still fits. Measured on our bench. 96GB of VRAM, $8,565 at launch.
- 1079.9 images/min on SAM ViT-Base
- 96GB, clears the SAM ViT-Base floor
- Rentable by the hour rather than bought
- 600W board rating
- Datacenter or workstation hardware, not a retail purchase
Segmentation cuts an image into objects: select the person, the product, the car, at pixel level. Meta's Segment Anything (SAM) is the standard open model for it and a building block in editing, labeling and robotics pipelines.
We measured 2 models for image segmentation with SAM on 11 GPUs. Speed is images segmented 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.
SAM ViT-Base: images/min by GPU
SAM ViT-Huge: images/min by GPU
Which models fit which card, for image segmentation with SAM
| Model | VRAM used | 8GB card | 12GB card | 16GB card | 24GB card | 32GB card | Licence |
|---|---|---|---|---|---|---|---|
| SAM ViT-Base | 4.7GB | Yes | Yes | Yes | Yes | Yes | Apache-2.0 |
| SAM ViT-Huge | 6GB | 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.
Every GPU x every image segmentation with SAM model (images/min)
| GPU | SAM Base | SAM Huge |
|---|---|---|
| NVIDIA B300 | 1719.6 | 399.4 |
| NVIDIA B200 | 1662.3 | 388.6 |
| NVIDIA H200 | 1469.3 | 325.2 |
| NVIDIA H100 80GB HBM3 | 1431.3 | 320.6 |
| NVIDIA RTX PRO 6000 Blackwell Workstation Edition | 1079.9 | 193.5 |
| NVIDIA L40S | 998.8 | 238.1 |
| NVIDIA A100 80GB SXM4 | 744.8 | 145.7 |
| NVIDIA A100 40GB SXM4 | 696.8 | 146.8 |
| NVIDIA A10G | 398.7 | 84.22 |
| NVIDIA L4 | 315.9 | 70.46 |
| NVIDIA T4 | 180.7 | 34.91 |
— = not measured on that card yet.
What the numbers show.
SAM ViT-Base: fastest on the NVIDIA B300 at 1719.6 images/min, 9.51x the slowest card we measured (NVIDIA T4); it used about 4.7GB of VRAM.
SAM ViT-Huge: fastest on the NVIDIA B300 at 399.4 images/min, 11.44x the slowest card we measured (NVIDIA T4); it used about 6GB of VRAM.
Which model to pick. On the same card, the NVIDIA B300, SAM ViT-Base runs at 1719.6 images/min in about 4.7GB; SAM ViT-Huge runs at 399.4 images/min in about 6GB. SAM ViT-Base gets through the work 4.3x as fast as SAM ViT-Huge, so the model you choose moves the speed as much as the card does.
How it compares. L40S: SAM ViT-Base 998.8 images/min, Depth Anything V2 Small 970.6, Depth Anything V2 Large 959.3, BiRefNet 863.8, SDXL Turbo 423.5 (3B). SAM ViT-Base beats all 4 here.
Cost on a rented GPU. 1,000 images of SAM ViT-Base: $0.011 on a A100 40GB SXM4 ($0.47/hr, 1 min), $0.067 on a B300 ($6.94/hr, 1 min, 6.0x the cost).
SAM ViT-Base: cost per 1,000 images on rented GPUs
| GPU | Cheapest rate | Speed (images/min) | Cost per 1,000 images |
|---|---|---|---|
| NVIDIA A100 40GB SXM4 | $0.47/hr | 696.8 | $0.011 |
| NVIDIA T4 | $0.14/hr | 180.7 | $0.013 |
| NVIDIA L40S | $0.79/hr | 998.8 | $0.013 |
| NVIDIA RTX PRO 6000 Blackwell Workstation Edition | $1.08/hr | 1079.9 | $0.017 |
| NVIDIA A100 80GB SXM4 | $0.95/hr | 744.8 | $0.021 |
| NVIDIA L4 | $0.44/hr | 315.9 | $0.023 |
| NVIDIA H100 80GB HBM3 | $2.14/hr | 1431.3 | $0.025 |
| NVIDIA H200 | $3.59/hr | 1469.3 | $0.041 |
| NVIDIA B200 | $5.98/hr | 1662.3 | $0.060 |
| NVIDIA B300 | $6.94/hr | 1719.6 | $0.067 |
Cheapest hourly rate we track on RunPod and Vast.ai, divided by the measured speed. Startup time and storage are extra.
Speed tiers for SAM ViT-Base. 30+ images/min: 11 (B300, B200, H200). 30 images/min means two seconds or less per picture.
Time per image. SAM ViT-Base: 0.0s per image on the B300, 0.3s on the T4. A batch of 100 takes 0 min on the fastest card and 1 min on the slowest.
VRAM for SAM ViT-Base. Measured peak 4.4GB, so 8GB is the smallest common card size; smallest card it ran on: T4 (16GB).
Power on SAM ViT-Base. Most efficient: L40S, 93W, 1.5 Wh per 1,000 images. Hungriest: B200, 204W, 2.0 Wh.
Our verdict
For image segmentation with SAM, 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 image segmentation with SAM?
- In our runs, the NVIDIA B300 at 1719.6 images/min on SAM ViT-Base. We measured 2 models on 11 GPUs for this page.
- How much VRAM do I need for image segmentation with SAM?
- The lightest model here, SAM ViT-Base, used about 4.7GB. 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 SAM ViT-Base?
- About 5GB. Smallest card that ran it: NVIDIA T4 (16GB).
- How much does it cost to run SAM ViT-Base in the cloud?
- $0.011 per 1,000 images on a NVIDIA A100 40GB SXM4 at $0.47/hr, cheapest of 10 rentable cards we measured.
- Can I run SAM ViT-Base on a 12GB, 16GB or 24GB card?
- It used 4.4GB at the precision we tested. 12GB: yes; 16GB: yes; 24GB: yes.
- Is the H100 80GB HBM3 or the A100 80GB SXM4 faster for SAM ViT-Base?
- The H100 80GB HBM3: 1431.3 vs 744.8 images/min, 92% 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.