Skip to content

PLATFORM / EDGE COMPUTE + VISION OS

Photon-3 Edge Compute + PhotonOS: One Stack for Industrial Vision Cells.

A purpose-built compute unit paired with a vision operating system engineered for brownfield lines. Drop a new automation cell into an existing line in under 72 hours — no fenced cages, no 18-month integration cycle, no PhD on staff.

Book a Discovery Walkthrough View Hardware Specs
  • 8 msEdge inference latency
  • 11 daysAvg. deployment-to-first-part
  • 14 imgZero-shot SKU training set
  • 72 hrCell integration target

01 / HARDWARE

Photon-3 Hardware Specs

Engineered for shop-floor reality: dust, vibration, 40°C ambient, and the I/O vocabulary of legacy PLCs. Photon-3 is a sealed, IP67 edge node with the safety and aerospace-grade quality certifications the rest of the industry splits across two SKUs.

  1. 01

    Compute & Inference

    SoC
    Strambotix Photon-3 SoM, 12-core ARM v9 + dedicated NPU
    Onboard NPU
    32 TOPS INT8, transformer-optimized
    Inference Latency
    <8 ms end-to-end vision pipeline
    Memory
    32 GB LPDDR5X ECC, 1 TB NVMe (industrial temp)
    OS
    PhotonOS 4.2 (read-only root, A/B partitions)
  2. 02

    I/O & Fieldbus

    Camera Inputs
    4× PoE+ GigE Vision (2.5G), 2× USB3.2, 2× CoaXPress-2.0
    Trigger Sync
    Hardware-stamped encoder, ±1 µs
    Industrial Bus
    EtherCAT, PROFINET, EtherNet/IP, CC-Link IE
    Discrete I/O
    8× isolated DI, 8× isolated DO (24V)
    Network
    2× 2.5 GbE, Wi-Fi 6E, 5G NR (option)
  3. 03

    Safety & Compliance

    Functional Safety
    ISO 13849-1 PL d, Cat 3
    Quality System
    AS9100D (aerospace) on the same platform SKU
    EMC / Environmental
    EN 61000-6-2, IP67 sealed, −20°C to +55°C
    Cybersecurity
    IEC 62443-3-3 SL2, signed firmware, TPM 2.0
    E-Stop
    Dual-channel hardwired + safe bus
  4. 04

    Mechanical & Power

    Form Factor
    260 × 180 × 65 mm, die-cast aluminum, fanless
    Mounting
    DIN-rail, panel, or VESA 75; M12 connectors throughout
    Power Input
    24 VDC ±20%, 35 W typical / 60 W peak
    Vibration
    IEC 60068-2-6, 5 g RMS, 10–500 Hz
    Serviceability
    Field-swappable SoM, 5-minute MTTR

02 / SOFTWARE

Inside PhotonOS

Four layers, one binary. PhotonOS ships as a sealed image on every Photon-3 — runtime, perception, the Zero-Shot engine, and fleet management, all reachable through the same SDKs and OT-friendly interfaces.

L1

Runtime & Hardware Abstraction

A real-time, containerized Linux with deterministic scheduling for vision pipelines. PhotonOS exposes a stable HAL so any camera, encoder, or robot controller registers as a first-class citizen — no vendor lock-in to a specific GigE or CoaXPress stack.

  • PREEMPT_RT kernel, 1 kHz tick, hardware-timestamped triggers
  • Containerd runtime with deterministic GPU/NPU pinning
  • Signed image updates, A/B partitions, atomic rollback
  • OPC UA server built in, Modbus adapter optional
L2

Perception Pipeline

A modular graph of stereo matching, photometric fusion, pose estimation, and 6-DoF grasp synthesis — composable per cell, executable entirely on the Photon-3 NPU. Drop in your own ONNX model at any node without forking the stack.

  • Calibrated stereo + structured-light fusion
  • 6-DoF pose estimation on reflective and deformable parts
  • Cycle budgets enforced per pipeline stage
  • Telemetry out via MQTT, Kafka, or gRPC stream
L3

Zero-Shot Engine & Fleet Manager

The patented Zero-Shot Part Recognition engine sits above the perception pipeline and handles new-SKU onboarding in minutes. The fleet layer above it gives your ops team a single pane for staging, deploying, and rolling back cell configs across 1,000+ sites.

  • Train on 14 reference images — no labeled dataset, no GPU cluster
  • Canary rollouts, signed cell configs, audit log per cell
  • MSSC-certified operator training curriculum bundled
  • SOC 2-ready telemetry, no customer imagery leaves the edge
Photon-3 edge compute unit mounted on a robotic arm with status LEDs lit
Photon-3, sealed in an IP67 enclosure. Same SKU on a John Deere weld cell and a Toyota Material Handling palletizing cell.

03 / DIFFERENTIATOR

Zero-Shot Part Recognition: Train on 14 Images, Run on Shift One.

Traditional machine vision wants thousands of labeled images, weeks of training compute, and a data engineer on call. Strambotix's patented Zero-Shot Part Recognition engine collapses that to a single shift: snap 14 reference images of a new SKU on the line, confirm the bounding box, and the cell is picking production parts before the next break.

  • 14 reference images. No labeled dataset, no GPU cluster, no overnight training job.
  • Operator-grade UX. A floor lead — not a vision engineer — completes onboarding.
  • Reflective, deformable, and mixed-tolerance parts. Validated across 340+ deployments 2021–2025.
  • Measured first-pass pick accuracy. 93.4% mean across customer cells audited by the Association for Advancing Automation.
Talk to a Solutions Engineer

04 / DEVELOPER ENTRY POINTS

Developer Entry Points

Photonos speaks every language your integration team already uses. Pick the surface that matches your stack — none of them are wrappers around a single proprietary API.

ROS 2

Native ROS 2 Node

Drop-in strambotix_bringup package. Topics for pose, pick points, and cell state map 1:1 to standard ROS 2 messages. Foxy, Humble, Iron, Jazzy supported.

Distribution · Humble LTS

gRPC

gRPC Streaming API

Bidirectional streams for pose, telemetry, and cell commands. Proto files are versioned and published; breaking changes follow a 12-month deprecation window.

Schema · v4.2

REST

REST Control Plane

OpenAPI 3.1 surface for cell provisioning, SKU enrollment, and rollback. Token-authenticated, idempotent, and OT-segmentable.

Auth · OAuth2 / mTLS

SDK

Python & C++ SDKs

Typed clients for both Python 3.11+ and C++17. ONNX import is first-class: bring a model from any framework, slot it into the perception graph at any node.

Wheel · strambotix==4.2.1

05 / AUDITED NUMBERS

Measured in Production, Not Slides.

Every figure below was captured from customer production cells and audited by a named third party. No synthetic demos, no cherry-picked lab runs.

EDGE INFERENCE LATENCY

<8ms

End-to-end vision pipeline on the Photon-3 onboard NPU, measured across customer cells.

FIRST-PASS PICK ACCURACY

93.4%

Mean across customer production cells audited by A3 (Association for Advancing Automation).

DEPLOYMENT-TO-FIRST-PART

11days

Average across the last 200 deployments, measured from signed SOW to first production part.

INTEGRATION-TIME REDUCTION

−71%

Versus traditional integrator timelines, measured across 340+ deployments from 2021 to 2025.

Recognized by G2 as a High Performer in Industrial Automation for four consecutive quarters (Q1 2024 – Q4 2024). Featured in Automation World's 2024 cover story on brownfield robotics and Fast Company's 2024 "Next Big Things in Tech" list.