--- title: Compute description: The compute control surface at --- # Compute ## Purpose COMPUTE is the operator's live control surface for GPU pods: what is running, what each pod serves, what it costs, and the cloud pricing underneath. All pod mutations happen here through PAT-authed step-up modals (scope `compute:write`). Deep link: `ops.useargus.co/#/compute` (nav group COMPUTE). ## What it shows * A filter toolbar for the pricing catalog: PROVIDER chips (ALL, AWS, RUNPOD), KIND chips (ALL, ONDEMAND, SPOT, COMMUNITY, SECURE), a SHOWING STALE / HIDING STALE toggle, and an EXPORT CSV button. * RUNNING PODS: one card per worker, running pods first. Each card shows a status badge, GPU model, dollars per hour, camera count (current/capacity), uptime, cost so far this run, an auto-off countdown when a TTL is set, a pinned marker, the RunPod pod id, and the stores and cameras it serves (an idle line when no cameras are assigned). Clicking a card opens the pod drawer. * ALL WORKERS: the full worker table (GPU, status, dollars per hour, cameras, pinned, drift, copy-ready argusctl commands, action buttons) plus the live RunPod cross-check. Orphan pods (live on RunPod with no worker row) render in a warning block; RUNNING orphans are flagged as billing now and untracked. The cross-check requires `RUNPOD_API_KEY` server-side. * MONTHLY SPEND: month-to-date GPU cost, current hourly burn, and breakdowns by provider and by tenant. When running orphans exist, a warning notes the burn figure excludes them and is understated. * PRICING REFERENCE: the GPU catalog (feed health per source, per-SKU rates) with a LAUNCH A GPU button. ## Actions * STOP, RESTART, TERMINATE on an active pod; RESUME and TERMINATE on a stopped pod. Each opens a step-up modal requiring a reason (at least 3 characters); RESUME and RESTART also take a "Keep up for (hours)" TTL (1 to 168, default 7, blank for no auto-stop). TERMINATE warns it permanently deletes the pod and cannot be undone. * LAUNCH A GPU opens a searchable SKU picker over the catalog. Picking a SKU opens the LAUNCH GPU step-up modal: GPU type, image, region, GPU count, volume, TTL, a high-cost acknowledgement checkbox, a required reason, and typing LAUNCH to confirm. Holding Alt while picking copies the equivalent `argusctl compute launch` command instead. * In the pod drawer, an orphan pod gets a TERMINATE THIS POD button (step-up, audited), the only in-console way to kill untracked spend. * All compute writes are per-operator rate limited (429 shows a retry countdown) and audited. ## Matching CLI ```bash argusctl compute list argusctl compute served argusctl compute up --gpu-type "NVIDIA GeForce RTX 4090" --image my/image:tag --reason "bring up inference" argusctl compute stop --reason "spin down idle" argusctl compute resume --reason "resume for demo" --ttl-hours 6 argusctl compute terminate --reason "tear down" --yes argusctl compute catalog --provider aws --rate-kind spot argusctl cost month ``` See [argusctl compute](/docs/cli/compute), [argusctl cost](/docs/cli/cost), and the [launch](/docs/compute/launch) and [stop, resume, terminate](/docs/compute/stop-resume-terminate) guides. ## Notes * Polled on the 15 second timer. Running-pod cards are patched cell by cell, so action buttons survive ticks; the catalog, workers, and monthly panels fetch independently so the slow catalog never blocks the pods panel. * In production today the served list reads idle until cameras are assigned to workers. * Monthly spend reads near zero until cost snapshots accrue. * Audit caveat: with no operator token set, the running-pods and workers panels render blank, which can read as "no pods". --- For a semantic overview of all documentation, see [/sitemap.md](/sitemap.md) For an index of all available documentation, see [/llms.txt](/llms.txt) For agent-facing discovery, including API and MCP surfaces, see [/agents.md](/agents.md)