Untether AI has collaborated with Arm to enable Untether’s AI inference accelerators to work with the latest generation of Arm automotive enhanced (AE) technology for autonomous vehicles (AVs) and advanced driver assistance systems (ADAS).
The trend for AI in AVs and ADAS is towards domain-specific architectures and heterogeneous compute optimized for AI and software-defined vehicles, Bob Beachler, VP product at Untether, told EE Times.
“What we’re seeing is that a lot of our automotive customers are trying to figure out ways to differentiate their hardware and make it better than an off the shelf system from Mobileye, Qualcomm or Nvidia,” Beachler said. “With the growing amount of AI in automotive infrastructure, they are looking at how to solve this problem with scalable solutions that go from level 2+ all the way to full autonomy.”
This type of system would require a scalable CPU architecture with the option for multiple AI accelerators, he said, which will be provided by the new Arm-Untether partnership.
Arm launched new AE cores last month, including a brand-new, server-class Neoverse CPU with automotive features, the Arm Neoverse V3AE. This core offers high single-thread performance with automotive functional safety features like transient fault protection. The IP provider is also working on extending its compute subsystem capability for chiplets via the AMBA CHI protocol over UCIe.
The aim of the new partnership is to provide plug-and-play compatibility between Neoverse-class Arm CPUs like the V3AE and Untether’ existing AI accelerator cards. The first step, Beachler said, has been to port Untether’s runtime to the Arm architecture in the lab to demonstrate Arm compatibility. Future work will ensure compatibility for Untether’s future chiplet-based accelerators, which will need to be CHI compatible. This will ensure Untether chiplets can be designed in alongside OEMs’ own Arm-based CPU chiplets in future ADAS and infotainment designs.
Automotive has been a focus for Untether since the company’s inception, Beachler said. The Canadian startup has General Motors as an investor and is currently working with automotive customers in Europe and North America for 2028-2030 models.
Untether is planning a separate automotive product line with all the necessary functional safety features. This product line will be in the form of chiplets and is due around the 2026 timeframe.
Untether’s RISC-V based AI inference architecture focuses on latency and energy efficiency, both crucial for automotive.
“At-memory compute gives us more TOPS per Watt,” Beachler said. “In the case of electrified vehicles, this goes to mileage. Are you going to sacrifice 10 or 15 miles of range on your electric vehicle using an inefficient AI accelerator?”
Untether’s AI accelerator cards for the data center are based on its RISC-V at-memory
compute chips. (Source: Untether)
The move towards chiplets will also help energy efficiency since higher levels of integration mean no need for PCIe, which uses plenty of power.
“The nice thing about our architecture, because it’s spatial, is we can right-size the devices for automotive,” Beachler said. “We won’t take a big data center class chip and shove it into automotive, we’ll right size it for the workloads that are needed and try to hit that sweet spot for throughput, latency and energy efficiency.”
ADAS perception systems like Bird’s Eye View might be running five or six neural networks simultaneously, he said. Flexibility to run different types of network is therefore crucial as the industry moves towards vision transformer (ViT)-based models like BEVFormer. Large language models will find automotive use cases in infotainment, Beachler added.
The next step for Untether will be to develop its chiplets over the next two years, paying careful attention to CHI compatibility and UCIe implementation, he said.
