GPU

Rent a GPU

GPU capacity reached over SSH, with an hourly price. Use it to run models, fine-tune, or anything else you'd do on a GPU server.

GPUs are rentals#

A GPU rental runs until you end it, Crowther support ends it, or the machine stops at our hardware partner. The order then shows Ended, when it ended and why. A monitoring agent going silent alone does not end your rental.

We email you and your workspace billing address every day after the first 24 hours while your rental runs. End it from your order when you no longer need it.

Ending is permanent

Ending deletes the machine and everything on its disk. It can't be recovered. Copy anything you need elsewhere before confirming termination.

What you get#

GPUChoose an available offer by model, GPU count and GPU memory
IsolationCheck the details panel's Type: a dedicated machine or a dedicated container. A container does not give you the whole host.
AccessUse the SSH command shown on your order, including its username and port
PriceCards show an hourly GPU price plus a 20% service charge, with the total hourly price

Rule: 30 minutes minimum

Once a GPU is running, it can't be terminated until it has run for at least 30 minutes. The Terminate button stays disabled until then.

1. Create an SSH key (once per computer)#

An SSH key is two files: a private key that stays on your computer and is never shared, and a public key (ending in .pub) that you give us. If you already have one, skip to step 2.

  1. Open a terminal

    Windows: press the Windows key, type PowerShell, open it. Mac: press Cmd + Space, type Terminal, open it.

  2. Create the key

    terminal
    ssh-keygen -t ed25519 -C "you@example.com"

    Press Enter to keep the standard place. Type a passphrase you'll remember (nothing appears as you type).

  3. Show the public key

    Windows (PowerShell)
    Get-Content $HOME\.ssh\id_ed25519.pub
    Mac / Linux
    cat ~/.ssh/id_ed25519.pub

    Copy the whole line that starts with ssh-ed25519. That's what you paste into Crowther. Never paste the file without .pub.

2. Order a GPU#

Opening View details checks that exact GPU's availability. You'll see Checking availability, then the result. If it is no longer available, ordering is disabled and similar alternatives appear; opening one checks it too. A Price updated notice shows the new total before you order. If we couldn't check, you can still order. Availability is checked again when you order; a price increase needs your confirmation.

High-reliability offers appear first. A Standard reliability tag marks a host with a shorter track record; it may be interrupted more often. Checks may reuse an answer from the last 30 seconds and do not reserve a machine.

  1. Open the catalog

    Choose GPUs in the sidebar (or New → GPU → Rent a GPU). Each card shows the GPU, its memory, the machine's RAM, where it is, and the price per hour. Use the filters to search, pick a GPU model, GPU count and region, or set minimum GPU memory, vCPUs and RAM and a maximum hourly price.

  2. Order, with your key

    An Admin or Owner presses Use this GPU on an available card, pastes the public key from step 1, and presses Order and start. Crowther checks the key as you paste it and shows its fingerprint.

  3. It starts

    The order starts without an approval step and shows under Your GPU orders on the same page. Open its name for the order details. Copy the connection command from the dashboard when it appears. We also email your workspace when it's ready; the dashboard's command is the one to use, because it includes the SSH port.

A workspace can have up to 3 GPUs starting or running at once. Terminate one to order another.

What each status means

StatusMeaning
StartingThe machine is being set up. This usually takes a few minutes.
RunningReady. The command to connect is on the order, with how long it has been running.
Didn't startSetup didn't finish.
TerminatedYou terminated it; the machine and its files are gone.

3. Connect#

Copy the command from your order under Your GPU orders. It looks like this:

terminal
ssh -p <port> <user>@<host>

It works the same in PowerShell, cmd, Mac Terminal and Linux: nothing to install. The first time, SSH asks whether to trust the machine. Answer yes.

Check your GPU is there: nvidia-smi shows it, with its memory and what's using it.

  • Leave Crowther's monitoring agent installed: it's how we keep an eye on the GPU for you.

More ways to connect#

A nickname, and VS Code

Add a block like this to ~/.ssh/config (on Windows, C:\Users\<you>\.ssh\config; create the file if it doesn't exist), with your machine's address:

~/.ssh/config
Host my-gpu
  HostName <host from your order>
  User <user from your order>
  Port <port from your order>

Now ssh my-gpu is enough. In VS Code, install the Remote - SSH extension, press F1, choose Remote-SSH: Connect to Host… and pick my-gpu.

Copying files

Use scp (built in everywhere) or rsync (Mac and Linux, better for big folders):

terminal
scp data.zip my-gpu:~/
scp my-gpu:~/results.csv .
rsync -avP ./dataset/ my-gpu:~/dataset/

Reaching a port on the machine

To use something running on the machine (a model server on 8000, Jupyter on 8888) from your own computer, forward the port over SSH and keep that window open:

terminal
ssh -L 8000:localhost:8000 my-gpu

Then open http://localhost:8000 on your computer. Nothing is exposed to the internet this way.

4. Serve a model#

With vLLM, any model that fits in your GPU's memory becomes an OpenAI-compatible API. Pick a model size that suits the GPU memory you chose:

on the machine
pip install vllm
export HF_HOME=~/hf-cache

tmux new -s model          # keeps it running after you log out
vllm serve <org>/<model> \
  --max-model-len 32768 \
  --enable-prefix-caching \
  --api-key "$MY_KEY" --port 8000
  • Leave tmux with Ctrl+B then D; come back with tmux attach -t model.
  • Reach it over SSH (see above): ssh -L 8000:localhost:8000 my-gpu, then call http://localhost:8000/v1.

5. Terminate it when you're done#

Press Terminate on the order under Your GPU orders. The machine and everything on it are deleted, so copy off anything you need first (see copying files).

Not before 30 minutes

A GPU can't be terminated until it has run for at least 30 minutes. Until then the button is disabled, and the order says when it will be available.

Monitoring#

Open the order's name under Your GPU orders to see its monitoring readings.

  • GPU utilisation, GPU memory, temperature and power, plus CPU, RAM and disk readings where available.
  • Charts offer 1 hour, 24 hours and 7 days. Metrics refresh while the page is open.
  • The page shows the monitoring connection and last report, or says it is waiting for the first report. Reload the page to refresh the order's status and last-report badge.

Please don't remove it

If the agent stops reporting, our team is alerted and may contact you.

Troubleshooting#

Permission denied (publickey)

SSH is using a different key from the one you ordered with. Compare ssh-keygen -lf ~/.ssh/id_ed25519.pub with the fingerprint shown when you ordered, or add -i with your key file to the dashboard's connection command: ssh -i ~/.ssh/id_ed25519 -p <port> <user>@<host>.

Connection refused or timed out

If the order says Running, wait a minute and try again: the machine may still be finishing its start. If it keeps failing, contact us.

Out of GPU memory

Lower --max-model-len, use a smaller or quantised model (FP8, AWQ), or check nothing else is still running with nvidia-smi.

My process stopped when I logged out

Run long jobs inside tmux (see above) so they keep running after you disconnect.