Table of contents

GPUs — podstack gpu

Rent and manage on-demand GPU instances (TrainPods) from your terminal.

podstack gpu types list                                        # available GPU types
podstack gpu pricing --gpu-type h100_sxm                       # price for a type
podstack gpu keys list                                         # your SSH keys
podstack gpu instances create --type h100_sxm --count 1 --tier on_demand
podstack gpu instances ssh <id>                                # connect

Discover GPUs and pricing

podstack gpu types list                     # list GPU types
podstack gpu pricing --gpu-type h100_sxm    # show pricing for a GPU type

GPU type names are lowercase with a variant/VRAM suffix. Current types include:

a100_40gb  a100_80gb  a30  a40  b200
h100_pcie  h100_sxm  h100_nvl  h200_sxm  h200_nvl
l4  l40  l40s  v100
rtx_3090  rtx_4090  rtx_5090  rtx_a4000  rtx_a6000
rtx_4000_ada  rtx_6000_ada  rtx_pro_6000

Run podstack gpu types list for the live set. podstack gpu pricing returns the price per hour, region, availability, tier (on_demand/spot), and whether instant-boot is available — for example, a100_40gb in india-south was ~₹120/hr on-demand at the time of writing.

SSH keys

You connect to instances over SSH, so register a public key first:

podstack gpu keys list               # list registered keys
podstack gpu keys --help             # add/remove keys

Launch an instance

The fastest way is the interactive picker:

podstack gpu launch                  # pick a GPU and launch it

Or create one explicitly:

podstack gpu instances create --type h100 --count 1 --tier on_demand

Key options: --type (GPU type), --count (number of GPUs), --tier (e.g. on_demand). Run podstack gpu instances create --help for the full list.

Manage instances

podstack gpu instances list          # your instances and their state
podstack gpu instances ssh <id>      # SSH into an instance
podstack gpu instances --help        # start/stop/terminate and more

Billing

GPU instances bill from your wallet while running. Terminate an instance to stop billing. See Wallet & pricing and the TrainPod section for the end-to-end provision → SSH → train workflow, including video walkthroughs.

  • TrainPods — the full on-demand GPU product and SSH tutorials.
  • Fine-tuning — managed training without managing a box.
  • Send & receive — move data to and from an instance.