Ambarella boosts auto AI silicon offering

  • January 17, 2022
  • William Payne

AI vision silicon company Ambarella has launched a family of AI domain controllers for automotive applications. The CVflow family of SoCs provides up to 500 eTOPs, a 30x boost in CPU performance over its previous generation.

Featuring up to 16 Arm Cortex-A78AE CPU cores, the CV3 family is designed to enable centralised, single-chip processing for multi-sensor perception—including high-resolution vision, radar, ultrasonic and lidar—as well as deep fusion for multiple sensor modalities and AV path planning.

The CV3 introduces Ambarella’s next-generation CVflow architecture, which incorporates the company’s development of an on-chip neural vector processor (NVP), with up to 500 eTOPS of AI compute, improved power efficiency and support for neural network (NN) inferences. The NVP supports advanced radar perception software, such as the Oculii adaptive AI software algorithms. A new floating-point general vector processor (GVP) offloads classical computer vision and radar processing from the NVP engines, and floating-point-intensive algorithms from the Arm CPUs.

The family of processors scales low-power SoCs for regulatory-class, forward-facing cameras to premium SoCs for Level 4 fully automated driving. It also supports both central and zonal architectures. In addition to covering the whole AD stack, the CV3 can simultaneously process in-cabin sensing applications, including driver and occupant monitoring. Ambarella’s complete line of AI perception SoCs is uniquely positioned to support the full range of viewing, recording, sensing and path planning applications, from basic ADAS to L4 AVs. This means that automakers no longer need to develop different software stacks for their entry-level, mid-range and premium vehicles, as they can use Ambarella’s unified CVflow platform across all models, saving engineering costs while enabling faster responses to market trends.

The CV3 also integrates Ambarella’s next-generation ISP, extending the company’s lead in image signal processing quality. By simultaneously supporting up to 12 physical or 20 virtual cameras, a single CV3 can process the entire sensor suite, which for typical L2+ deployments includes 10 cameras, five radar modules and numerous ultrasonic sensors. Additionally, higher-performance stereo and dense optical flow engines provide greater depth and motion perception.

“With our new CV3 AI domain controller family, we are now capable of running the full ADAS and AD stack with a single chip, while providing unprecedented performance and power efficiency,” said Fermi Wang, President and CEO of Ambarella. “Through enhancements to our on-chip ISP, along with simultaneous radar processing that can take advantage of our Oculii adaptive AI algorithms, Ambarella is helping the automotive industry unlock greater levels of perception accuracy across all environmental conditions to realise the promise of autonomous driving.”