FAQs on NVIDIA® Jetson Thor™

Jetson AGX Thor is NVIDIA's next-generation AI computing module, delivering up to 2070 FP4 TFLOPs with the Blackwell GPU architecture, 2560 CUDA cores, 96 Tensor cores, and 128 GB LPDDR5X memory. It is designed as the ultimate platform for humanoid robotics and real-time edge AI.

e-con Systems provides Thor-compatible cameras and robost ECU solutions, enabling multi-sensor support, multiple interfaces, high-resolution imaging, real-time imaging. With Holoscan Sensor Bridge (HSB) integration, e-con's cameras achieve ultra-low-latency streaming directly into GPU memory, ideal for AI vision workloads.

Yes. In addition to Lattice's HSB, e-con Systems offers 10G compact HSB camera solutions, which include its own Holoscan Sensor Bridge board and its proprietary TintE® ISP designed to integrate with Jetson Thor seamlessly.

The 10G compact HSB is a space-efficient, production-ready solution that supports:

• a variety of camera modules with different supported resolutions of up to 20MP
• Ultra-low-latency image streaming using e-con's FPGA-based TintE ISP
• Camera-over-Ethernet (CoE) with direct GPU memory transfer
• Flexible ISP pipelines (Userspace ISP, NVIDIA Argus ISP, or TintE ISP)

TintE ISP from e-con Systems is an image signal processor built on FPGA technology, optimized for embedded vision platforms. It provides real-time 4K @ 60 fps processing, advanced automatic imaging controls (focus, exposure, white balance), and a finely tunable imaging pipeline, all deployable on major FPGA architectures. Unlike traditional ISP pipelines, TintE ISP offers:

• Customizable image tuning for different sensors and use cases
• Flexible pipeline options (NVIDIA Argus ISP, or TintE ISP)
• Scalable performance and high-resolution inputs

The key benefits of the e-con HSB board are:

• Small form factor
• User Configurable GPIO Connector
• Onboard TintE ISP
• Control over ISP Parameter via Holoscan SDK
• Flexible Cabling
• Passive cooling for Ethernet PHY

e-con Systems offers a wide range of NVIDIA Thor-compatible cameras across key interfaces as follows:

• USB Cameras - rapid prototyping and evaluation
• RouteCAM GigE Cameras - ONVIF-compliant Ethernet vision solutions
• DepthVista_HSB - 3D Time of Flight (ToF) 10G Holoscan Camera
• Small form factor 10G Holoscan Camera Solutions - Camera over Ethernet (CoE) with ultra-low latency and different resolutions.

e-con's robust ECU compute platform is optimized for NVIDIA Thor and enables real-time AI processing at the edge. It supports AMRs (Autonomous Mobile Robots), humanoid robots, and AI-powered machines, ensuring fast decision-making in dynamic environments.

You can connect the QSFP 25GbE port of the HSB using an RJ45 Ethernet cable along with the following adapters:

• RJ45 to SFP+ adapter
• SFP+ to QSFP adapter

The HSB board communicates using UDP packet-based Ethernet communication with the host system, which enables reliable, low-latency streaming of high-resolution, multi-camera data directly to GPU memory.

No. Daisy chaining is not supported in the HSB. But using an FPGA, the data can be aggregated from multiple sensors and sent the stream to the host via Ethernet with lower latency.

Yes. Multiple HSB boards with their respective cameras can be connected through a 10GbE or 25GbE Ethernet switch to the Jetson Thor platform. This enables multi-camera setups for robotics, medical imaging, and industrial applications.

Yes. It supports multiple cameras with hardware customization, with limitations of up to 10 GbE data.

Yes. Resolution switching and camera feature controls (such as exposure, gain, and frame rate adjustments) can be performed at runtime through the e-con HSB interface, offering developers greater flexibility during deployment.

No, HSB uses its own layer with UDP packetization.

e-con's HSB cameras support the following:

• Argus ISP is suitable for standardized NVIDIA workflows.
• TintE ISP (FPGA-based) provides ultra-low-latency, customizable image processing, ideal for time-critical applications like robotics and medical imaging.
• Userspace ISP - can integrate their own ISP

If the camera is not detected, the HSB software layer will throw a NOACK (No Acknowledgment) error. In such cases, users should verify the hardware setup and connections (power, Ethernet/connector seating) using the procedures mentioned in the NVIDIA website Installation: AGX Orin system running Ubuntu or L4T with Linux Sockets - NVIDIA Docs

The e-con's HSB includes a user-configurable GPIO connector, enabling interfaces like UART, I²C, SPI, and optional connections to IMU, GPS, RADAR, or other sensors—ideal for multi-modal perception systems.

The Jetson Thor provides industry-leading video processing performance for multi-camera, high-resolution embedded vision systems. Its capabilities for H.265/H.264 video encoding and decoding are as follows:

• Video Encoding : It can encode up to 6x 4Kp60 or 12x 4Kp30 streams. For high-density applications, it handles up to 24x 1080p60, 50x 1080p30, or 48x 1080p30 (H.264) streams.
• Video Decoding : The device can decode up to 4x 8Kp30 streams, up to 10x 4Kp60, or 22x 4Kp30 streams. For lower resolutions, it supports up to 46x 1080p60 or 92x 1080p30 streams.

The Jetson Thor provides industry-leading video processing performance for multi-camera, high-resolution embedded vision systems. Its capabilities for H.265/H.264 video encoding and decoding are as follows:

25GbE Ethernet is backward compatible with 10Gbps cameras. e-con Systems' Thor-compatible cameras can be seamlessly connected with Thor series modules.

The cameras support a wide range of sensors, including IMX715, IMX568, AR0234, IMX900, AR2020, ISX031 and Depth sensor (ToF sensor).

The CUDA cores accelerate parallel image/video processing (de-noising, stitching, scaling), while Tensor cores accelerate AI inference (object detection, recognition, segmentation). Together, they enable multi-camera, real-time AI vision systems.

The ability to partition Thor's GPU into multiple distinct instances is made possible by the Multi-Instance GPU (MIG). For cameras, this means one GPU can simultaneously handle different AI workloads (e.g., one instance for obstacle detection, another for driver monitoring, another for medical imaging) without interference.

The ability to partition Thor's GPU into multiple distinct instances is made possible by the Multi-Instance GPU (MIG). For cameras, this means one GPU can simultaneously handle different AI workloads (e.g., one instance for obstacle detection, another for driver monitoring, another for medical imaging) without interference.

Yes. e-con Systems provides custom sensors, firmware/software customization, HSB board and end-to-end integration support, making it easy to adapt NVIDIA Thor solutions to specific AI vision cases.

e-con's NVIDIA Thor-compatible cameras and ECUs are ideal for:

Humanoid robotics - real-time motion & perception

Autonomous mobile robots (AMRs) - navigation & obstacle detection

Medical imaging - high-resolution, real-time visualization

Industrial automation - anomaly detection & quality inspection

Automotive systems - cabin monitoring & driver assistance

You can purchase the Jetson AGX Thor Developer Kit directly from the NVIDIA website.