FrameBuff

Docs / AI & GPU

Lend your GPU

Lend some of a machine's tools to someone else for a set time. For example: atlas makes images for Mum every night from 22:00 to 06:00.

What you lend

You can lend to a person or a group you added in People (family, friends or work), or to one of your own paired machines. You choose which tools to lend, and the other side can run only those. Your files, cameras, vault and keys stay private.

The work runs on your machine, as jobs. You can end a loan at any time. Ending a loan deletes nothing.

Set up a loan

Open the lending page, under Life in the menu, and start a new loan. You choose:

  • What to lend: tools that do a job, such as image generation or chat with your models.
  • Who: a person, a group or one of your machines.
  • When: a window in half-hour steps, every day, on weekdays or at weekends. Or on request.
  • Until when the loan lasts.
  • The queue: owner first, first come first served, or one at a time.
  • A cool-down: minutes before the window ends when no new job starts.
  • For each tool and each person: ask first or use freely.

A loan to a person waits for their yes. A loan to one of your own machines starts at once. Each person also has a switch, so you can pause one member without ending the loan.

Lending your models

A loan can include chat with your machine's models. On the other machine, they appear in the model picker under your machine's name. You can also give the right to load and swap models, which is off by default. Only your own local models are lent, never an online model.

When you need the machine back

When several people want the machine at once, it serves them in this order:

  1. You, at the machine.
  2. Your own machines.
  3. Family and friends.
  4. Outside programs using an API key.

You count as using the machine when its keyboard or mouse was used in the last 5 minutes, when one of your own FrameBuff jobs is on the graphics card, or when the card is busy: over 50% for at least 4 seconds, or more than half its memory used by other programs.

Every loan states its rule up front. The default is Pause when I'm using it; up to 50% when I'm idle. You can also slow lent work to 10, 25 or 50% while you use the machine, and give it 25 to 100% while you are idle.

  • An answer that is already running is never cut off. It finishes first.
  • Work that is waiting starts by itself when you stop.
  • The other side sees why it waits, for example "paused while the owner's GPU is busy". They never see what you are running.

Limits and usage

Each person can have their own limits:

  • GPU-minutes a day or a week.
  • AI tokens or requests a day.
  • Storage on your machine, in GB.
  • A priority, your own hours when nothing of theirs starts, and which tools and models they may use.

Over the limit, their job either waits or is refused. You choose which, and the job always says why. Both of you see the same usage line, such as "used 12 of 30 GPU-minutes today".

A GPU-minute is a minute that their job's own processes held your card. When the card cannot tell which process uses it, the job's running time counts instead, and the usage line says so.

What you see of their work

Their prompts, files and results stay theirs. FrameBuff's pages, notifications and logs show you the tools and times of their jobs. They never show the arguments, logs or results. For their stored files you see sizes and counts, never names or content. A finished private job is cleared from your machine soon after its result is delivered.