Start here

Concepts

The words the rest of the docs use, each in a sentence or two.

The docs use a handful of terms. Most of them borrow from a forge because that is what the building is.

Forge

A site with GPUs in it. F1 in Reykjanes is live. F2 in Québec is next. Each forge has its own queue, storage and API id (f1).

One machine with 8 GPUs on one NVLink fabric. F1 has two, called node A and node B.

One container run on 1 to 16 GPUs, under one plan, with a time limit. Jobs are the only unit we bill.

Stage	What happens	Billed
`queue`	Waiting until every GPU the job asked for is free at once.	No
`heat`	Image pulled, GPUs cleared, storage mounted.	No
`strike`	Your container is running.	Yes, per second
`quench`	Container exited. Scratch and GPU memory wiped, invoice line written.	No

When a job leaves quench it has a result: succeeded, failed, stopped, preempted or hardware.

How a job is paid for: strike, bank, anvil or ember. See Plans.

Prepaid GPU hours, bought in blocks of 1,000. Jobs on the bank plan draw from it.

A whole node rented to one team for a month or more. Jobs on it start at once and nobody else runs there.

Cheap, preemptible work that runs on GPUs nobody is using. It gets 30 seconds of notice before it stops.

Our name for the telemetry feed: load, memory, temperature and power for every GPU, once a second. See Telemetry.

The monthly public record of every GPU hour in a forge: what ran, what sat cold, and every minute of downtime. See Heat ledger.