GPU list - Compute

compute gpu list
compute gpu list --json

That output is the source of truth for names and list ¢/hr. Marketing pages can lag.

v0.1 SKUs

Name	Also accepted	Provider	Role
`H100-SXM`	`H100`, `H100-80GB`	RunPod Secure	Reserved public lead
`MI300X`	—	Hot Aisle	AMD; stock can be tight
`RTX-3090`	`3090`, `RTX3090`	Vast.ai interruptible	Lowest locked rate, 24 GB
`RTX-4090`	`4090`, `RTX4090`	Vast.ai interruptible	Common 24 GB training card
`RTX-5090`	`5090`, `RTX5090`	Vast.ai interruptible	32 GB consumer
`RTX-6000-ADA`	—	Vast.ai interruptible	48 GB workstation
`L4`	—	Vast.ai interruptible	24 GB inference
`A100-40GB`	`A100-40`	Vast.ai interruptible	40 GB datacenter
`A100-80GB`	`A100-80`	Vast.ai interruptible	80 GB datacenter
`H100-SXM`	—	Vast.ai interruptible	Same name as RunPod; qualify it

H100-SXM exists on both RunPod and Vast.ai. A bare --gpu H100-SXM is ambiguous — use --provider runpod or --provider vastai, or write vastai/H100-SXM. Vast.ai SKUs are interruptible. The host can reclaim the machine. Compute then ends the run and confirms teardown so you are not left paying for a paused box.

Decorator vs CLI

@app.function(gpu="H100-80GB", image=compute.Image.cuda_pytorch(), timeout=600)

compute run simple_mlp.py::train --gpu H100-SXM
compute run simple_mlp.py::train --provider vastai --gpu RTX-4090

The CLI --gpu / --provider flags win when both decorator and CLI are set. simple_mlp.py is in First run. To choose from compatible GPUs using workload analysis:

compute run simple_mlp.py::train --gpu auto
compute run simple_mlp.py::train --gpu fastest
compute run simple_mlp.py::train --gpu cheapest

auto balances predicted execution time and provider cost. fastest and cheapest optimize one of those estimates directly. See Run for the quote and confirmation flow.

Images

NVIDIA (RunPod and Vast.ai): compute.Image.cuda_pytorch()
AMD: compute.Image.rocm_pytorch()

Use the image that matches the SKU. Mixing CUDA images onto MI300X (or the reverse) is a failed run, not a silent fallback. Rates: Pricing. Caps: Limits.