GPU tier selection

Choose the right GPU for your workload and budget.

Available GPUs

GPU

VRAM

Rate

Per-second

RTX 4090

24 GB

$0.50/hr

$0.000139

A100 80GB

80 GB

$2.10/hr

$0.000583

H100 SXM

80 GB

$3.80/hr

$0.001056

When to use each tier

RTX 4090

Best for most tasks. Inference with models under 20GB, image generation, fine-tuning small models, data processing, and general CUDA workloads. Start here unless you know you need more VRAM.

A100 80GB

Large model inference (70B parameter models in 4-bit), training runs that need more than 24GB VRAM, multi-batch processing, and workloads that benefit from higher memory bandwidth.

H100 SXM

Maximum performance. FP8 training, large-scale inference, workloads sensitive to memory bandwidth, and anything where execution speed matters more than cost.

VRAM requirements for common models

Model

VRAM needed

Min GPU

SDXL

~7 GB

4090

Llama 3 8B (fp16)

~16 GB

4090

Llama 3 70B (4-bit)

~40 GB

A100

Llama 3 70B (fp16)

~140 GB

H100 (multi)

Whisper Large V3

~4 GB

4090

FLUX.1 Dev

~12 GB

4090

Auto-selection rules

If you don't specify a GPU tier, Volspan defaults to the RTX 4090. Your agent can choose the appropriate tier based on the workload. There is no automatic upsizing — if your code runs out of VRAM on a 4090, the run will fail with an OOM error rather than silently upgrading to a more expensive GPU.

We recommend letting your agent select the tier. Most agents will pick the right GPU based on the model size and task description.

Need help? Reach out at [email protected]