Skip to content

How-to: check live GPU stock across all datacentres before booking a pod

This is a worked task that spans a few commands. The reference pages (rp doc stock gpu, rp doc stock dc, rp doc pod create) document the individual flags; this page shows them composed into a pre-booking check.

Goal: confirm the GPU type you want is available somewhere, see which datacentres have GPU stock, and book a pod pinned to a datacentre — or let the scheduler choose one for you.

Steps

  1. Find which GPU types are in stock. rp stock gpu lists one row per GPU type with a STOCK column — the fleet-wide availability for the product and cloud you asked about. Filter to pod-relevant hardware and a secure cloud, and require at least one free on a host:

$ rp stock gpu --product POD --min-count 1 --cloud SECURE

The ID column holds the display name rp pod create --gpu takes — names contain spaces, so quote them. --product defaults to POD,SERVERLESS; use --product SERVERLESS for serverless availability, or --min-cuda 12.1 to filter by CUDA version.

  1. See which datacentres have GPU stock. rp stock dc prints a table keyed on the DATACENTER id, with REGION and GPUS (how many GPU types have any stock there, not how many cards), plus the S3_API, GLOBAL_NETWORK, NETWORK_VOLUME_TYPES, and COMPLIANCE columns:

$ rp stock dc

The DATACENTER value is exactly what rp pod create --dc takes.

  1. Get the per-datacentre, per-GPU-type breakdown. rp stock dc --json prints the datacentre records including the gpuAvailability array, which lists each GPU type and its availability for every location. This is the source of truth for "which GPU types are in stock where" — the datacentre dimension is already carried here, so no second query or client-side join is needed:

$ rp stock dc --json | jq '.[].gpuAvailability'

  1. Book the pod. Pass the chosen type and datacentre:

$ rp pod create --name dev --image runpod/pytorch:2.2.0 \ --gpu "NVIDIA L4" --dc EU-RO-1

Omit --dc and the scheduler places the pod wherever that GPU type is free.

Notes

  • rp stock gpu is the aggregate view by construction: its STOCK column is per-GPU-type availability across the fleet. There is no --dc or --all flag on rp stock gpu, and no option that sums the rp stock dc rows into a fleet total — use rp stock gpu directly rather than summing datacentre rows. (rp stock cpus --dc does filter by datacentre, but that is a different verb.)
  • There is no rp command that prints one flat cross-DC-per-GPU table, but rp stock dc --json already carries the datacentre dimension inside gpuAvailability, so a single call answers "which GPU types are in stock where" — and rp stock gpu's STOCK column alone cannot tell you where a type is free.
  • For a network-volume-backed pod, the datacentre is fixed by the volume. Use step 3 to confirm the GPU type you want is in stock in that specific DC before booking.
  • rp stock gpu also takes column-level filters (all apply to both the table and --json): --cloud SECURE|COMMUNITY keeps only types offered on that tier, --vram-gb N / --vram N keeps types with at least N GB of VRAM, --stock NONE|LOW|MEDIUM|HIGH keeps a single availability level, and --cuda <ver> keeps types with that CUDA version available. --sort <column> reorders rows by ID, DISPLAY, VRAM_GB, CLOUD, SECURE_PRICE, COMMUNITY_PRICE, STOCK or CUDA.