Skip to main content

Compute Profiles

The reusable shape of compute — GPU vendor, count, instance type, autoscaling bounds.

Compute Profiles

Compute profiles answer one question for every other Workbench tab: where does this run? A profile bundles GPU type, GPU count per node, instance class, and autoscaling bounds — the same name then shows up in Sessions, Endpoints, Training Jobs, and Sweeps.

Read-only by design

Compute profiles are managed by workspace admins, not end users — they affect cost, capacity, and quotas. The Workbench UI is intentionally read-only here. To add or change a profile, talk to your admin or use the Vantage admin console.

info

Need a profile that doesn't exist? Common requests: an H100×8 profile for big training, a CPU-burst profile for data prep, a regional profile near a specific dataset. Open an admin ticket with the use case and the rough hours/month estimate.

Picking the right profile

WorkloadSuggested profile shape
Notebook + light experimentation1× GPU (T4 / L4) or CPU-large; min=0 so it scales to zero when idle.
Single-node training run1× A100/H100 with whatever max your budget allows.
Distributed trainingMulti-GPU profile (4× or 8× per node), node count set in the training-job sizing.
LLM inferenceProfile sized to fit the model; H100×1 for 7-13B, H100×4 for 70B.
Heavy data prep / ETLCPU-large with high memory; no GPU.

See Reference for every field on a compute profile.

⌘I