ZF Hikes Performance of ProAI Supercompute Platform
ZF’s next-gen hyper-adaptable central controller prepped for 2024 production.
Although multinational supplier ZF is widely recognized for conventional components such as its transmissions and braking systems, at April’s Auto Shanghai – one of the few major auto shows to take place in the past year as a consequence of COVID-19 – ZF introduced a next-generation automotive-grade supercomputer, the ProAI system, a new variant of the controller that it initially introduced in 2019 at CES.
A recent conversation with Dr. Martin Fischer – a member of the ZF board of management whose extensive portfolio includes Electronics and ADAS, Passive Safety Systems, Active Safety Systems North America Region, South America Region Corporate Quality – focused on the “software-defined vehicle” and the role that hardware plays in it. Specifically, he said that ZF’s latest ProAI controller ticks all the boxes in terms of what is needed for handling the needs for SAE Level 2 to Level 5 automated driving, as well as a wide range of other vehicle-management responsibilities. The new ProAI, he said, is flexible, scalable and powerful.
As a supplier with global scale, ZF is certainly familiar with the interests and requirements of its OEM customers. But the development of automated driving and ADAS capabilities is ongoing, so everyone involved is continually learning. To that end, ZF is obtaining firsthand knowledge of the requirements for automated driving. It acquired a company named “2getthere,” which operates autonomous shuttles in Europe. An electric vehicle in the 2getthere fleet can accommodate up to 22 occupants and measures 6 x 2.1 x 2.8 meters (19.7 x 6.9 x 9.1 feet).
The 2getthere shuttles have been operating in dedicated traffic lanes, using markers for guidance. “We will be launching fully autonomous vehicles into mixed traffic,” Fischer said, adding that to do so, they need “the highest levels of computational power.” The 2getthere project will be a working shuttle service in Germany, not an engineering exercise, running routes in both Mannheim and Friedrichshafen (ZF’s hometown).
This, in addition to OEM demand, led the company to develop the new ProAI platform, as well as middleware and applications that run on it, thereby having a fully capable hardware and software offering. The latest ProAI, Fischer says, can accommodate a variety of systems on a chip (SoC) from companies including Nvidia, Qualcomm and TI.
The ProAI unit, which is in a commonized 240 x 138 x 49-mm housing to accommodate scale, is capable of handling SoCs with compute capacity from 20 to 1,000 trillion operations per second (TOPS); the 1,000-TOPS SoC is the recently introduced Nvidia Atlan, which Nvidia says will be ready for use in 2025 model-year vehicles.
Because of this compute-capacity range, Fischer says that they have had to address the issue of power consumption and heat dissipation. He cites the Nvidia Orin SoC, which has a 250-TOPS capability, as an example: Orin requires 80 watts in operation. “Think of a conventional lightbulb operating at 80 Watts. It gets pretty hot.”
So ZF has developed three approaches to cooling the ProAI platform. There is passive cooling (aka, “the good old heat sink,” as Fischer puts it), air cooling (two fans, one drawing in air and the other exhausting it) and liquid cooling. “This is what people want to avoid,” Fischer said, but concedes that at the higher operational level when more power is involved, the liquid-cooling layout is required. Regardless of the thermal-management layout, ProAI’s footprint is the same.
Notably, ProAI has been engineered so that while ZF has been able to increase the overall performance of the unit by as much as 66%, power consumption has been reduced by up to 70% – or about 3 TOPS per watt, on average.
A specific vehicle’s features and controls architecture largely would determine the number of ProAI controllers installed. Fischer said this essentially is a decision made by the OEMs and is predicated on the performance of the SoC “package.” It could be possible to run on one unit operating as the ECU for an ADAS or automated-driving vehicle, but Fischer points out the way control functions are partitioned within the vehicle architecture is the determining factor. He also noted that there are considerations of redundancy and cybersecurity.