Use case · computer vision
Computer vision GPUs: detection budgets measured in epochs
Vision workloads are bursty by nature — a training sprint, a giant batch job, then nothing. Per-second billing fits perfectly: a YOLO11 epoch for about $0.02, an archive sweep for the price of coffee, zero idle spend between sprints.
The vision cards, ranked
| Tier | GPU | VRAM | On-demand | Interruptible | Why this card | Action |
|---|---|---|---|---|---|---|
| Good | RTX 3090 | 24 GB | $0.104 | $0.052 | 24 GB at rock-bottom pricing — big batches for augmentation-heavy training. | Deploy |
| Better | RTX A5000 | 24 GB | $0.161 | $0.080 | ECC + blower cooling for week-long training queues; the reliability pick. | Deploy |
| Best | RTX 4090 | 24 GB | $0.262 | $0.131 | Fastest epochs per dollar — the default for YOLO, segmentation and ViT fine-tunes. | Deploy |
Batch-inference fleets: Tesla T4 from $0.104/hr — TensorRT detectors barely notice the smaller card.
Two typical jobs, costed
| Train YOLO11m, 100 epochs | 50k images @ 640px, RTX 4090 interruptible | ≈ $1.57 |
|---|---|---|
| Sweep 1M archived photos | TensorRT detector, 4 × Tesla T4 in parallel | ≈ $0.19 |
| Dataset volume 150 GB | images + labels + runs, one month | $12.00 |
Pipelines that scale down to zero
- Ultralytics/MMDetection in the PyTorch template — pip install and train.
- Export to TensorRT before batch runs — 3–5× throughput on the same card.
- GPU video decode — NVDEC keeps CPU out of the hot path for camera streams.
- Weights on a volume, instances disposable — retrain Fridays, pay Fridays only.
Computer vision GPUs: FAQ
Which GPU for training YOLO models?
A RTX 4090 trains YOLO11m on a 50k-image dataset at roughly 7 minutes per epoch — about $0.02/epoch interruptible. 24 GB fits big batch sizes at 640px; step up to multi-GPU only past a few hundred thousand images.
What about batch inference over an archive?
Cheap cards shine: a Tesla T4 at $0.104/hr pushes hundreds of frames/s with a TensorRT-exported detector. A million images costs a few dollars — spread the queue over several instances and it finishes over lunch.
Do you support video pipelines?
Yes — NVDEC/NVENC are exposed in containers and VMs, so decode → detect → encode runs entirely on GPU (DeepStream, PyAV, ffmpeg builds in the PyTorch template).