Ruggedizing sensor-fusion infrastructure for AVs
Progress toward full vehicle autonomy depends on the ability to real-time process large volumes of sensor information. With Elon Musk’s comments on the future of LiDAR at Tesla’s Autonomy Investor Day in April, the AV (autonomous vehicle) implementation and economics debate continues to be a popular topic.
Whether the sensor fusion is camera- or LiDAR-driven, we expect the compute load to remain roughly similar. Considering a self-driving car generates four terabytes of stored data in about an hour and a half—a number that will grow as capabilities evolve—manufacturers need to rely on sensor fusion infrastructure that is not only powerful and versatile, but also hardened to withstand real-life road conditions. Most of our customers have settled on 48-56 Intel Xeon cores and approximately 360 TFLOPS (trillion floating-point operations per second) of tensor-core capability within an AV system.
An average electronics enclosure may be suitable for stationary, temperature-controlled office environments, but it is not adequate for mobile and dynamic conditions in which autonomous vehicles operate.
Just like our brain‘s synaptic characteristics for vision, hearing, and feel, a sensor fusion unit is the critical core of an autonomous car’s situational awareness. If a sensor-fusion unit is not immediately responsive, that car can pose a very real threat to its driver and everyone else on the road. The AV platform provides a plethora of challenges including 1500 W of 12-V DC power, excessive heat dissipation, and limited cooling capabilities. The processing electronics in an AV system need to be protected against operational and environmental elements.
Automated driving systems should be tailored specifically to the environmental conditions in which the systems are expected to operate. Manufacturers need to ensure the compute/sensor platform doesn’t freeze or overheat in extreme weather and can reliably operate in temperatures ranging from -40°C to +55°C (-40°F to +130°F). As an example, circuit board topography needs to be protected from humidity with a layer of conformal coating. Consistency and reliability are enhanced by using an automated process.
The most impressive compute power, data-handling capabilities and storage capacity are not enough, if not stabilized in a rugged enclosure. The case enveloping critical autonomous vehicle infrastructure has to withstand the harshest road conditions, including potholes or collisions.
Ensuring the structural integrity of a server starts with designing the chassis from relatively thick and subsequently stiff aluminum plates. This technique enhances reinforcement of the front, rear, and side plates. Bonding the base and mechanically fastening it to the side walls increases the torsional stiffness of the enclosure. At the same time, servers need to be optimized for SWaP (size, weight, and power) to fit mobile environments without impacting performance. This is accomplished using a modular architecture which can easily be adapted to accommodate the size of the vehicle.
As one of the lightest, strongest, and corrosion resistant metals, aluminum is a perfect material to protect sensor-fusion servers. The chassis’ strain-hardened aluminum construction helps limit weight and improves thermal conductivity. A box-in-box design increases the resonant frequency of the chassis, limiting deflection in vibration, lowering solder joint fatigue on the electronics.
Balancing the SWaP and unprecedented processing requirements necessitates reliable thermal management. Standard conductive and convective heat-transfer mechanisms need to be employed to extend the thermal limits of commercial off-the-shelf motherboards. A cooling system with low-noise internal fans and increased air flow is key to maximizing compute density while maintaining safe electronics operating temperatures. Liquid cooling is frequently employed to enhance the thermal capabilities of the COTS-based system.
CPU and GPU manufacturers innovate and update technology almost as fast as the demand for processing power grows. With an average lifespan of a chip at 18 months, a data sensor-fusion system has to be upgraded with newer technology shortly after it is deployed. This makes it critical for autonomous vehicle designers to rely on an infrastructure foundation that easily accepts change and upgrades. Design modularity plays a role in future-proofing the platform.
When it comes to autonomous vehicles, failure is not an option. Manufacturers need to work with partners that understand technology strengths as well as physical vulnerabilities to anticipate the inevitable impact of the operational environment. Ruggedizing autonomous vehicle systems, just like cybersecurity, cannot be an afterthought. It needs to be considered early in the design to ensure reliability over a long operational life in environments unforgiving to common computer and electronics devices.
When lives are on the line, physical integrity of a sensor fusion unit is as critical as its compute capabilities.