Your machines · GPU box · twenty minutes

Add a GPU box to your workspace

A tower with a real graphics card is the machine that thinks next to your robots and your data: a vision model at the cameras' frame rate, a policy for the arm, a language model the agent can pick — with nothing leaving the building. It joins like the Pi; the card is what it brings.

A compact GPU workstation tower on a wooden workbench, the graphics card visible through the glass side

What you need

  • A PC with an NVIDIA card. 24 GB runs a 27B-class model; 32 GB (RTX 5090) the same with room for context; 96 GB (RTX PRO 6000) a 70B at FP8 or a 120B mixture-of-experts, with the KV cache an agent's long conversations need.
  • Ubuntu 22.04 or 24.04 with the NVIDIA driver installed — nvidia-smi shows the card.
  • Power, a network, SSH or a keyboard for the first minute.
  • No card yet? Rent one for the afternoon. It pairs the same way and is gone when you leave.

Install and pair

The same command as on the Pi. The node notices the GPU and puts it on the tile.

$ curl -fsSL https://daslab.run/node | sh
installed daslab 1.1.63 for x86_64
service enabled and started

  pairing code  H4RN-7PWC
  open          daslab.run/pair

waiting to be paired…

Type the code at daslab.run/pair, pick the workspace, and the tile appears with gpu on it.

paired with workspace labhaus as lab-gpu
offers:     x86_64 · 16 cores · 192.168.1.0/24 · gpu nvidia · camera
ready

Run a scene on it

In a scene, tap Add and pick lab-gpu under Machines, or say run this scene on lab-gpu. From the next job on, the scene's computer is the box: the card, the cameras on it, the lab's network. Things that want a GPU next to the bench:

Run the detector over the last hour of camera frames at full rate and chart the counts.
Fine-tune on today's episodes and tell me the loss curve when it is done.
Start the sim and roll out two hundred episodes with the new policy.
Embed every note in this scene and find the ones about the gripper.

Serve a model to your scenes

Start a model server on the box (vLLM, SGLang, Ollama, anything that speaks the OpenAI API) and publish its port:

$ vllm serve Qwen/Qwen3.8-27B-FP8 --served-model-name qwen3.8-27b --max-model-len 131072 --host 127.0.0.1
INFO  Uvicorn running on http://127.0.0.1:8000
$ daslab node publish models 127.0.0.1:8000

In a scene, add a cell, Model on a machine, with the machine lab-gpu and the service models. The model shows in that scene's picker as qwen3.8-27b on lab-gpu, beside the ones we host. The agent uses it like any model, through the connection the node already holds open, with no port on your router and no public address, and the price is what the box costs you: nothing per token. More on your own model.

What fits

CardMemoryRuns well
RTX 4090 / 509024 / 32 GB27B-class at 8-bit, with room for a long context on the 5090
RTX PRO 6000 (Blackwell)96 GB70B at FP8, GPT-OSS-120B, Qwen3.8-27B at 262k context, several agents at once
Two cards48–192 GBMore at once; one bigger card beats two smaller ones for a single model
Rented H100 / L40S80 / 48 GBAnything above, by the hour: rent one

When the box is off

Scenes pinned to it say so and wait. A model it serves is not in the picker until it is back.

Take it out

daslab node leave, or delete the tile.

The other machines