Deployment sizing
We size disks, servers and network from how many devices capture at once and how long you keep the data. The result is an estimate; contact us for a quote or delivery.
Workstation and cluster run in your machine room, off the public internet. Video, labels and models stay on site. SaaS is the exception.
Deployment
Workstation and cluster stay in your machine room. Data does not go to the public internet.
Setup
Capture devices use a 1 GbE port, then connect to the workstation. Capture and labelling can share a subnet.
Estimated use
Count two kinds of device together: cameras on the robot, and cameras worn on a person’s head or body. Add up how many are capturing at the same time.
Robots capturing now
Head-worn cameras (a person doing the task)
Devices in total
10People using the platform at once
How long to keep data (months)
Model training (optional)
GPU memory
Trainable base models
- π₀ / π₀.₅
- Isaac GR00T N1.x
- SmolVLA
- Spirit-v1.5
- Wall-OSS
- OpenVLA
- EO-1
GPU count
Recommended GPU
RTX PRO 5000 Blackwell (48 GB)Robot capture
Usually three 480p compressed colour streams (MJPEG). Joint data is counted separately.
480p compressed colour (MJPEG)
Capture hours per day
Capture days per year
Extra color images
Depth images
What to buy
Single-machine install (Docker Compose 2.20+)
1 × Tower workstation
CPU / memory
32 cores / 128 GB
GPU
1 × RTX PRO 5000 Blackwell (48 GB)
Reference model
Dell Pro Precision 9 T6 or equivalent
| Role | Qty | CPU / memory | GPU | Reference model |
|---|---|---|---|---|
| Tower workstation | 1 | 32 cores / 128 GB | 1 × RTX PRO 5000 Blackwell (48 GB) | Dell Pro Precision 9 T6 or equivalent |
Storage
8 × 32 TB · Erasure coding 10+4 (up to 4 disks can fail) · 146 TB
Drive bays to leave free (three-year capture)
9
Facility
Floor-standing; no rack required. Place it in a machine room or equipment room. Peak 1.8 kW
Sizing uses compressed recordings: about 6 Mbps per 480p colour stream, about 3.5 Mbps per 720p video stream. A typical robot (three 480p cameras) is about 8.5 GB per hour; a typical headset (four 720p cameras) is about 6.9 GB per hour. These rates follow common camera layouts and bitrates in open datasets on Hugging Face, such as LeRobot and ALOHA. Uncompressed or several HD streams are an order of magnitude larger — encode on the device.
The system keeps the original recording and also writes a training export, so capacity is counted at 1.17×. Usable space then subtracts two copies or erasure coding, and assumes you expand disks at about 80% full.
How many devices can capture at once depends on the network and servers. How long you can keep data depends on disks. Size them separately. Wire the floor; do not use Wi-Fi as the main path.
Want this configuration?
Send us the estimate. An IO-AI delivery engineer will check your site and available machines and send a quote.