Three GPU tiers optimized for different workloads. Billed per second.
VRAM: 24 GB GDDR6X
Performance: 82 TFLOPS FP16
Price: $0.50/hr ($0.0083/min)
Best for inference, small fine-tuning jobs, and experimentation. Models up to ~20B parameters (quantized) fit comfortably. Fastest cold start of all tiers.
VRAM: 80 GB HBM2e
Performance: 312 TFLOPS FP16
Price: $2.10/hr ($0.035/min)
Best for large model inference, training runs, and batch processing. Handles 70B parameter models in FP16. High memory bandwidth for data-intensive workloads.
VRAM: 80 GB HBM3
Performance: 990 TFLOPS FP16
Price: $3.80/hr ($0.063/min)
Best for production training, large-scale inference, and latency-critical workloads. 3x faster than A100 for transformer architectures thanks to the Transformer Engine.
First run on a cold worker includes container pull and initialization. Subsequent runs on warm workers start near-instantly.
4090 — ~5s cold, <1s warm
A100 — ~8s cold, <1s warm
H100 — ~10s cold, <1s warm
GPUs are provisioned on-demand from RunPod's global fleet. 4090s have the highest availability. A100 and H100 may occasionally queue during peak demand — typical wait is under 30 seconds. Your vault is only charged once execution begins.
Need help? Reach out at [email protected]