Today, robots rely on a wide range of sensors, including cameras, radar, and LiDAR, to maintain balance, manipulate objects, and move around obstacles. But giving them more ways to sense their surroundings also creates a connectivity problem. Each sensor tends to use a different protocol or proprietary interface, such as MIPI CSI-2 and I2C, requiring increasingly complex point-to-point connections. As more sensors are embedded in robots as well as industrial, military, and medical systems, system costs increase, too.

Microchip Technology offers one approach to solving the problem with its Polarfire FPGA-based Ethernet sensor bridge. The device sits between the main system processor and a combination of high-bandwidth sensors such as cameras as well as low-bandwidth devices, e.g., inertial sensors. Instead of wiring each sensor directly to the processor, the sensor bridge brings together the data, pre-processes it, and then streams it to system memory over a single high-bandwidth Ethernet connection.

Microchip, which recently released Revision 2.0 of its FPGA Ethernet Sensor Bridge, is betting that this architecture can make sensor connectivity more secure, scalable, and future-ready. Designed for physical AI systems such as humanoid robots, the new board supports NVIDIA's Holoscan Sensor Bridge (HSB) technology while offering a 60% smaller footprint and the ability to handle twice as many cameras as the first-generation design.

It gives engineers an alternative to complex, multi-interface sensor connections by replacing them with a more scalable 10-Gb/s Ethernet architecture, reducing wiring, system cost, and overall complexity.

HSB: Bringing Sensor Processing Closer to the Edge

HSB is a sensor-over-Ethernet technology developed by NVIDIA to simplify real-time sensor processing in edge systems based on the company’s Jetson Thor and Orin series of computer modules. Instead of linking each sensor directly to the processor, HSB uses a small FPGA or high-performance MCU to pool all of the data before processing, packetizing, and synchronizing it (see figure). Using Ethernet, the HSB technology is designed to deliver data directly into GPU memory with under-a-millisecond latency.

By applying HSB, the sensor bridge can deliver sub-microsecond sensor synchronization while taking over sensor-processing tasks from the main processor. Processing these operations closer to the edge enables the Jetson to devote itself fully to high-compute AI inference with ultra-low latency and maximum overall throughput. Microchip said its second-generation sensor bridge features 2 GB of DDR4 x32 memory to help reduce the burden on the Jetson and 125 MB of SPI flash for FPGA configuration.

The software stack integrates Holoscan Sensor Bridge IP, which enables the sensor-to-Ethernet pipeline. As NVIDIA expands its efforts in physical AI for robots, Microchip isn’t alone in building FPGA-based sensor bridges compatible with Holoscan. For instance, Lattice released a reference board in collaboration with NVIDIA, combining its CertusPro-NX FPGA with HSB IP for real-time sensor acquisition and processing. Altera also introduced its Stratix 10 Sensor Processing Kit.

HSB offers other advantages over commonly used sensor interfaces, including longer-range connectivity. Camera interfaces such as MIPI and GSML are designed for relatively short connections. Other sensor interfaces like I2C, GPIO, and UART are suited to even shorter distances, spanning only a couple of centimeters. In contrast, HSB uses Ethernet to connect sensors and processors over meters of cable, giving engineers more flexibility when placing both.

Rather than relying on a fixed set of sensor lanes with predetermined speeds, HSB leverages a network architecture using high-bandwidth Ethernet links ranging from 10 up to 100 Gb/s, which in turn enables scalable, distributed multi-sensor systems not possible with traditional point-to-point links. The modular design means that engineers can handle higher-resolution cameras and additional sensors without redesigning the host’s dedicated sensor interfaces, according to NVIDIA.

HSB is also a potentially more scalable architecture. With standard point-to-point links, each additional camera or sensor generally requires a separate physical lane to the processor, which can rapidly consume the sensor interfaces in the host processor. With HSB technology, multiple cameras and sensors can instead connect through Ethernet switches, allowing systems to scale beyond the number of dedicated camera lanes or sensor ports on the host.

These benefits are a boon to industrial automation and robotics systems where power efficiency, small form factors, and deterministic latency are all key. NVIDIA said HSB is engineered with end-to-end safety protocols, meeting up to SIL 2.

HSB Sensor Bridge Runs Everything Through Ethernet

Microchip said its second-generation sensor bridge acts as the sensor connectivity center for the system. The new board supports sensor, video, and display interfaces, including MIPI CSI‑2 for cameras, HDMI and DisplayPort for video, as well as I2C, UART, and GPIO. It features a high-pin-count FMC connector that enables future sensor expansion without requiring a redesign when upgrading sensors or adding more of them.

The sensor bridge can connect up to four cameras through its four 4-lane MIPI CSI-2 D-PHY interfaces and integrates a camera connector that’s compatible with NVIDIA’s Jetson platform. The device also comes with hardware for determining the time delay between a camera capturing an image and the system processing it. Microchip said it can be used to validate end‑to‑end latency, which is a critical metric for safety-critical, physical AI systems such as humanoid robots that need to react to their surroundings in real-time.

The board uses the high-speed SerDes inside the FPGA to output images, video, and other sensor data through dual 10-Gb/s SFP (small-form-factor pluggable) ports, enabling a direct connection to 10G Ethernet networks, said Microchip.

“Developers want to spend their time building high-value edge AI applications, not stitching together proprietary sensor interfaces,” said Shakeel Peera, vice president of Microchip’s FPGA business unit. “With low-power PolarFire FPGA technology at its core, this second-generation Ethernet sensor bridge delivers a power-efficient, secure foundation in a significantly reduced form factor to help teams move faster from development to deployment in edge AI systems.”

The security and safety features inside its FPGAs assist in protecting edge AI devices and support reliable long‑term deployment. Integration with the NVIDIA Holoscan SDK helps accelerate development with optimized libraries, AI models, and reference applications.

The HSB-based board, USB‑C powered for streamlined rack deployment, is being offered at a lower price point than the first generation.