Altera Connects to Nvidia Holoscan Sensor Bridge with FPGAs, Targeting the Physical AI Market
Altera has launched FPGAs for the Nvidia Holoscan Sensor Bridge (HSB), targeting physical AI markets such as robotics and autonomous driving. The architecture connects in the order of sensors, FPGA, Ethernet, and GPU; the FPGA aggregates multi-protocol sensor data and handles low-latency tasks such as localization, motor control, and image preprocessing, then writes directly into Nvidia GPU memory over Ethernet, bypassing the CPU to reduce latency. Altera offers four FPGA series—Cyclone, Agilex 3, Agilex 5, and Agilex 7—and is working with Terasic and Critical Link to develop a sensor board and SOM ecosystem.
Body · AI translation
A
The body content of the article will be changed to this letter size.
Scroll position bar
Altera Director Lee Nam-hoon introduces the Holoscan Sensor Bridge. (Photo: THE ELEC)
Altera is targeting the physical artificial intelligence (AI) market, including robotics and autonomous vehicles, with field-programmable gate arrays (FPGAs) designed for Nvidia’s Holoscan Sensor Bridge (HSB).
Lee Nam-hoon, a director at Altera, discussed the technology at the “Semiconductors That Power Humanoids” conference held alongside the Semiconductor Exhibition (SEDEX) 2026 at COEX in Seoul on Oct. 16.
“The key to physical AI, including robots and humanoids, is interaction with the real world,” Lee said. “This requires numerous sensors, including cameras, lidar, radar and ultrasonic sensors.”
“We developed the Holoscan Sensor Bridge with Nvidia to connect these sensors to graphics processing unit (GPU) platforms,” he added.
The Holoscan Sensor Bridge uses an architecture that connects sensors, FPGAs, Ethernet and GPUs in sequence. Altera’s FPGAs and Nvidia’s GPUs divide processing tasks to improve computational efficiency.
The FPGA aggregates data from sensors using different communication protocols. It also handles latency-sensitive operations in real time, including positioning, motor control and image preprocessing.
The aggregated data is converted into Ethernet packets and transmitted directly to Nvidia GPU memory. The GPU then focuses on large-scale parallel computing and AI inference.
This architecture allows sensor data to bypass the central processing unit (CPU), reducing latency, according to Altera.
It also simplifies wiring. Instead of requiring separate cables for individual sensors, the system can use a single high-speed Ethernet connection.
Altera said the approach also reduces the limitations associated with the short transmission distances supported by MIPI (Mobile Industry Processor Interface), a camera connectivity standard.
Altera offers four FPGA families for the Holoscan Sensor Bridge: the entry-level Cyclone, Agilex 3 for compact robots, Agilex 5 for next-generation robots including humanoids, and the higher-end Agilex 7.
Developers can select chips according to requirements such as robot size, sensor count and data bandwidth.
Altera also supplies two modular development kits for Agilex 5-based reference designs.
The company is working with partners including Terasic and Critical Link to develop an ecosystem of dedicated sensor boards, including system-on-modules (SOMs).
An SOM is a compact circuit board containing key components such as an FPGA, memory and power circuitry. Robot developers can connect an SOM to a main board without having to design their own semiconductor chips.
JEONG IL JOO 1week@thelec.kr
My comment