Devices overview
Axelera AIPUs are the silicon. The Voyager SDK is the software that runs on them. Wingman is the agent that drives both — this tab takes you from an empty machine to pipelines running on your card.
Axelera builds AI accelerators in two generations, Metis and Europa, shipped as M.2 modules, PCIe cards, and a compute board. Hardware generations & product names maps every product to its generation. The Voyager SDK is the software stack around that hardware: a model zoo, a compiler that quantizes networks for the Axelera APU, and inference.py, the runner that executes pipelines on the card.
You do not have to learn that stack up front. In the desktop app, Agent Dev turns a plain-English ask — “build a people-detection pipeline for my USB camera” — into a validated, running pipeline. The pages in this tab teach both paths: ask the agent, and understand the commands it runs for you.
What a finished pipeline looks like
What you can build
- Detection: Find people, vehicles, or your own classes in video, cameras, and streams — YOLO-family models come precompiled in the zoo.
- Tracking: Follow objects across frames with stable IDs, including Re-ID trackers that survive occlusion and camera pans.
- Pose estimation: Keypoint detection for bodies in motion — run
yolov8npose-cocoon a webcam and see skeletons live. - Segmentation & more: Instance and semantic segmentation, plus depth estimation and classification, all from prebuilt zoo models.
Chat cannot see your card
What works with your setup
| Your setup | What works | Where to start |
|---|---|---|
| Linux machine with an Axelera card and the Voyager SDK | Everything: build, run, benchmark, and validate pipelines on real silicon. | Set up the Voyager SDK & your card, then the first detection tutorial. |
| Linux machine, no card yet | Agent Dev still builds the full pipeline and packages it with a RUN_HERE.sh runner you carry to a machine with an Axelera card. | The tutorials work the same — the run step happens on the target machine. Over ssh, Run on a device you cannot install Wingman on does that step for you. |
| Browser only (hosted Chat) | Planning, Voyager SDK questions, reviews, and team knowledge. No hardware access from the cloud. | Plan in Chat, then hand off to Agent Dev in the desktop app. |
The tutorial ladder
The tutorials in this tab form a ladder. Each rung assumes the one before it, and each has a video.
- 1 · First detection pipeline: The flagship first success: prompt Agent Dev, approve the plan, and watch detection run on your card — about 8 minutes.
- 2 · Cameras & streams: Point a working pipeline at USB cameras and RTSP streams, run several at once, and add tracking and pose.
- 3 · Model zoo: Find the right prebuilt model by task or framework and run it against files, cameras, streams, and datasets.
- 4 · Go further: Deploy your own models, benchmark variants to pick a winner, and port whole projects from other ecosystems.