Ambarella, Inc. announced during CES the latest additions to its CV3-AD family of automotive AI domain controllers—the CV3-AD635 and CV3-AD655 systems-on-chip (SoCs). The new CV3-AD635 supports a sensing suite that includes multiple cameras and radars to enable mainstream L2+ feature sets, such as highway autopilot and automated parking, in addition to meeting the GSR2 and NCAP standards. Additionally, the CV3-AD655 enables advanced L2+ (also called L2++) with urban autopilot, as well as support for additional cameras, radars, and other sensors. With the previously announced flagship CV3-AD685 SoC, which targets L3/L4 systems — along with the China-focused CV72AQ SoC—the CV3-AD family now covers the full range of AD and ADAS solutions, from mainstream to premium passenger vehicles. Additionally, this is the automotive industry’s broadest performance range for any software-compatible AI domain control family, spanning an 18x AI performance increase from the CV72AQ to the CV3-AD685.
All SoCs in the CV3-AD family feature Ambarella’s CVflow 3.0 AI accelerator. This highly efficient, proprietary AI engine combines neural network acceleration—including support for the latest NN architectures such as transformers and bird’s-eye-view (BEV) networks—along with general vector processing capabilities for traditional computer vision and efficient radar processing. The CVflow’s high AI processing performance and industry-leading power efficiency enable CV3-AD customers to design vehicles with rich feature sets based on safe and reliable AD and ADAS capabilities while reducing thermal-management cost and complexity. That significantly lower power consumption also allows customers to deploy electric vehicles with smaller batteries, reducing cost and weight; or to alternatively increase the EV range using the same battery. As with the rest of this family, the CV3-AD635 and CV3-AD655 are fabricated in Samsung’s advanced 5nm automotive process technology.
Ambarella’s CV3-AD SoC portfolio meets the processing requirements for automotive OEMs’ entire catalogs, from mainstream to premium models, by scaling the performance of each member’s compute engines. For example, the CV3-AD655 offers double, and CV3-AD685 six times, the NN processing performance of the CV3-AD635. Likewise, the CV3-AD635 integrates four Arm Cortex A78AE CPU cores and the CV3-AD655 eight cores. A dual-core, lockstep pair of Cortex-R52 CPUs is also included in both the CV3-AD635 and CV3-AD655. Additionally, these SoCs target ASIL-B at the chip level, via an ASIL-D safety island. Also, integrated on-chip is a GPU to render 3D surround views.
Easy portability across the entire CV3-AD family is achieved by sharing the same architectural approach, SDK, and tools.