NVIDIA Jetson AGX Thor
Embedded / edge
Validated Product Solution Brief · NVIDIA Holoscan Ecosystem
Arducam Holoscan Multi-Camera Solution
8x GMSL2 in. Dual 10G SFP+ out. Synchronized and low latency.
Arducam Holoscan Multi-Camera Solution bridges up to 8x GMSL2 vision sensors directly into the NVIDIA Holoscan ecosystem. Utilizing the NVIDIA Holoscan Sensor Bridge (HSB) protocol over Dual 10G SFP+ optical links, it delivers hardware-synchronized, millisecond-level latency, with a link reach of up to 15 m over GMSL coax and more than 100 m over 10G SFP+ optical links to the host. Engineered for Embodied AI, humanoid robotics, and high-concurrency perception systems, it offers native compatibility with NVIDIA Jetson AGX Thor, IGX, Jetson AGX Orin, and DGX Spark.
8× GMSL2 Vision Sensors Dual 10G SFP+ Optical Links Hardware Sync · gPTP / IEEE 802.1AS Jetson AGX Thor · IGX · DGX Spark
Scroll
8×GMSL2 Inputs
2× 10GSFP+ Outputs
20 GbpsRaw Bandwidth
µs-LevelHardware Sync
100+ mOptical Reach
01 · Why Holoscan
Why NVIDIA Holoscan
A full-stack platform that unifies sensor-to-GPU streaming, a modular software stack, edge-to-cloud scalability, and production-grade tooling for real-time edge AI.
What Holoscan + Sensor Bridge Deliver
Ultra-low latency
Deterministic pipelines for time-critical AI tasks.
Ultra-fast data handling
Sustain heavy multi-stream throughput for real-time decision making.
Sensor-to-GPU, by design
The Holoscan Sensor Bridge integrates diverse sensor data directly into the GPU for low-latency, synchronized, real-time AI processing.
Hardware-level sync
gPTP / IEEE 802.1AS time synchronization aligns all connected sensors for tightly synchronized, jitter-free data — a foundation for 3D SLAM and spatial AI.
Modular graph SDK
Python / C++ plug-and-play operators; low-code, reusable — saves months of low-level driver work.
Edge-to-cloud
One pipeline from Jetson → IGX → AGX / DGX, no code rewrite; built-in Triton and MONAI.
Multi-sensor compatibility
Supports a wide range of sensor types (MIPI, GMSL2/3, Ethernet) via software-defined configuration.
The Gap It Closes
Legacy MIPI-CSI CSI
PCB-level reach (~30 cm), point-to-point, custom bridge/FPGA to scale.
Legacy GMSL GMSL
Reliable for a few cameras, but scaling and vendor-proprietary PHY make multi-camera designs complex.
Both Shared limits
Sensor data hits CPU first → extra latency, OS jitter, CPU heat.
Other approaches Alternatives
General-purpose stacks and cloud-centric inference often add latency or lack a unified sensor-to-GPU pipeline.
Holoscan Arducam-ready
A unified platform combining all of the above — with guaranteed performance and faster time-to-market.
02 · Why Arducam
Why Arducam
Two capabilities set Arducam apart in the Holoscan ecosystem — everything else is table-stakes we simply meet.
From Reference to Product
NVIDIA proved the architecture.
Arducam makes it deployable.
Engineering the bridge, the carrier, and the SDK so you ship perception — not prototypes.
01
CoreDual 10G GMSL2 aggregation
Our core differentiator: 8× GMSL2 video inputs aggregated into dual 10G SFP+ optical outputs—delivering ~20 Gbps of raw bandwidth with hardware synchronization and 100m+ long-reach capability. High-bandwidth, extended-range GMSL2 aggregation is Arducam’s core strength, with a hardware design built to seamlessly scale to dual 25G SFP28 interfaces for next-generation workloads.
02
CoreSensor ecosystem
A broad, validated sensor portfolio (Sony, onsemi, OmniVision and more). Host-side I2C dynamic register configuration adds new sensors through software files — zero FPGA or hardware changes. You are not locked to a single sensor vendor.
Supporting capabilities — industry-standard, not unique to Arducam
03
Full-stack hardware design
FPGA + deserializer + carrier design; modular SoM/carrier decoupling enables fast, cost-effective customization.
04
Production-grade build
IP67, industrial temperature range, 15 m long-reach coax, hardware-level synchronization.
05
The result
An 8x GMSL2 + Dual 10G SFP+ HSB-compatible bridge that turns NVIDIA’s reference design into an engineered, productized, scalable solution.
03 · Demo
Live Demo — Arducam + Holoscan
Proof it runs: multi-camera perception, synchronized and low-latency, on the NVIDIA Holoscan stack.
Live on Holoscan
Demo — 8-channel GMSL camera live preview on Holoscan
Eight ISX031 cameras streamed concurrently through the Holoscan Sensor Bridge to a Jetson AGX Thor host — stable ~30 fps per channel, video output across all 8 cameras.
Demo library is being finalized — depth, SLAM and teleoperation recordings coming soon.
04 · Features
Key Features
01
High-Bandwidth Architecture
8x GMSL2 + Dual 10G SFP+, 20 Gbps raw bandwidth. Full-load multi-camera streams with no frame drops, and streaming offloaded from the host CPU.
02
Expansive Sensor Ecosystem
Out-of-the-box integration with mainstream sensors. Host-side I2C dynamic configuration adds new sensors via software files — no FPGA / hardware changes.
03
Ultra-Low Latency
Deterministic, time-sensitive streaming. FPGA hardware encapsulation and direct optical links offload streaming from the host CPU and software buffering for seamless teleoperation.
04
Dual-Group FSYNC
8 channels split into two 4-camera groups, externally triggered. Millisecond-to-microsecond alignment across all 8 cameras — jitter-free data for 3D SLAM and spatial AI.
05
Production-Ready SDK
Fully validated Linux drivers and native NVIDIA Holoscan SDK compatibility. Eliminates months of low-level driver development.
06
Modular Carrier Design
FPGA core SoM decoupled from the carrier board for fast, cost-effective customization (deserializer, interfaces, port density).
05 · Hardware & System Architecture
Hardware & System Architecture
Sensor-to-GPU by design: eight GMSL2 sensors deserialize into the bridge FPGA, packetized by Hololink IP and streamed over dual 10G SFP+ optical links straight into NVIDIA GPU memory.
On the Bridge
- Data preprocessing on FPGA
- Flexible interconnects and data packetization
- Reduce GPU overload
- Hardware encapsulation, CPU offload
On the Host
- Better task management and scheduling
- Low latency and better video streaming / compression
- Extensive libraries, pipelines, and frameworks
- Accelerates development — decision output in real time
06 · Cameras
Adapted Cameras & Specifications
Validated sensor portfolio — example list, to be finalized.
| Sensor (Camera) | Resolution | Key Features | Output Format | Modes (pixel · FPS) | Status |
|---|---|---|---|---|---|
| Sony ISX031Automotive-grade | 3 MP |
| YUV422 (UYVY) | 1920×1536 · 30 | Example |
| Sony IMX900Global shutter | 3 MP |
| RAW10 | 2048×1552 · 30 | Example |
| Sony IMX477Stacked BSI | 8 MP |
| RAW10 | 3840×2160 · 202016×1520 · 201920×1080 · 60 | Example |
| Sony IMX219Wide compatibility | 2 MP |
| RAW10 | 1920×1080 · 20 | Example |
| Onsemi AR0234Stereo | 2.3 MP×2 |
| YUV422 (UYVY) | 3840×1200 · 302560×720 · 301280×480 · 100 | Example |
| OmniVision OX08B40Automotive | 8 MP |
| YUV422 (UYVY) | 3840×2160 · 30 | Example |
New sensors are added via host-side I2C register files — no FPGA or hardware changes required.
07 · Platforms
Supported Platforms
Native compatibility across the NVIDIA edge-to-data-center range.
NVIDIA IGX Orin
Industrial edge
Compatible — HSB reference
NVIDIA Jetson AGX Orin
Embedded / edge · up to 275 TOPS
Compatible — HSB reference
NVIDIA DGX Spark
Desktop / data-center class
Compatible — HSB referenceIntegration layer: NVIDIA Holoscan Sensor Bridge (HSB) protocol over Dual 10G SFP+.
08 · Value
Value & Significance
A validated foundation for the industries building the next generation of perception.
Embodied AI & Humanoid Robotics
Multi-camera synchronized perception and low-latency teleoperation accelerate embodied-AI training and deployment.
Data Acquisition (EgoSync)
Long-reach 10G SFP+ optical links (up to 100 m) + hardware-synced multi-sensor capture enable large-scale, high-concurrency training-data collection and replay.
Surgical Robotics & Medical Imaging
Hardware-level synchronization and a production-grade, safety-conscious software stack support medical-grade real-time control and imaging.
Autonomous & Industrial Vision
8-channel 360° awareness, IP67, 15 m reach for rugged AMR, warehouses, and outdoor inspection.
One validated reference design — de-risking adoption across robotics, data capture, and medical perception.
Request Evaluation
Engineering samples and customized carrier board development available upon request for qualification.
Request Hardware Spec Sheet
Full datasheet, mechanical drawings and bill of materials for the 8x GMSL2 sensor bridge.
Request DocumentApply for DevKit Evaluation
Qualify the bridge with your sensors and Holoscan pipelines on Jetson AGX Thor or IGX.
Apply NowArducam · Holoscan Multi-Camera Solution
10 · FAQ
Frequently Asked Questions
Quick answers on the Holoscan Multi-Camera Solution — architecture, platforms, synchronization, and evaluation.
Q1What is the Arducam Holoscan Multi-Camera Solution?
A validated sensor bridge that connects up to 8× GMSL2 vision sensors into the NVIDIA Holoscan ecosystem over dual 10G SFP+ optical links. Built on the NVIDIA Holoscan Sensor Bridge (HSB) protocol, it delivers hardware-synchronized, millisecond-level latency streaming with the data path offloaded from the host CPU — turning NVIDIA’s reference architecture into a deployable, productized solution.
Q2How many cameras can it support, and what’s the bandwidth?
Up to 8× GMSL2 inputs feeding 2× 10G SFP+ outputs — roughly 20 Gbps of raw bandwidth. The eight channels are split into two 4-camera groups with dual-group FSYNC for microsecond-level alignment across all sensors, with no frame drops and no host CPU overhead.
Q3Which NVIDIA platforms are supported?
NVIDIA Jetson AGX Thor is fully Validated; IGX Orin, Jetson AGX Orin, and DGX Spark are supported via the HSB reference. The integration layer across all of them is the Holoscan Sensor Bridge protocol over dual 10G SFP+.
Q4What is the maximum cable and optical reach?
Up to 15 m over GMSL coax per camera, and more than 100 m over 10G SFP+ optical links to the host. That long reach makes it well suited to industrial, warehouse, and large-scale robotic deployments where the compute host sits far from the sensors.
Q5How is multi-camera synchronization achieved?
Hardware-level time sync via gPTP / IEEE 802.1AS aligns all connected sensors to microsecond-level precision. Combined with dual-group FSYNC, this eliminates jitter for 3D SLAM, spatial AI, and precise multi-view reconstruction.
Q6Which cameras and sensors are supported?
A growing portfolio including Sony ISX031, IMX900, IMX477, IMX219, onsemi AR0234, and OmniVision OX08B40. New sensors are added through host-side I2C register files — no FPGA or hardware changes required. (The full validated list is being finalized.)
Q7How do I request an evaluation or spec sheet?
Use the Request Evaluation options above — request the full hardware spec sheet, or apply for a DevKit to qualify the bridge with your own sensors and Holoscan pipelines on Jetson AGX Thor or IGX.
Simplifying Embedded Vision, Innovation Made Easy.
© Arducam · Holoscan Multi-Camera Solution Brief
