Original reporting: VentureBeat. Post date: January 6, 2020.
Ambarella Expands Into Robotics With CVflow-Based Platform
Santa Clara-based chip designer Ambarella announced a new robotics platform built on its CVflow architecture — the same low-power AI processing architecture the company has used in its computer vision system-on-chip (SoC) products. Alongside the platform announcement, Ambarella revealed a partnership with Amazon Web Services (AWS) to simplify machine learning model optimization for developers targeting CVflow-powered devices.
Both the robotics platform and the AWS integration were demonstrated at CES 2020 in Las Vegas.
Background: From Video Chips to AI Vision SoCs
Ambarella went public in 2011 and originally built its reputation making low-power chips for video cameras. Over time, it leveraged that imaging expertise to develop computer vision capabilities, eventually launching the CVflow architecture as a foundation for AI-accelerated, low-power vision processing. The CVflow SoC family includes the CV2, CV22, CV25, and S6LM chips, each targeting different performance and power envelopes.
The CVflow Robotics Platform
Target Applications
The new robotics platform provides a unified software infrastructure for robotics perception across the CVflow SoC family. It is designed for:
- Automated guided vehicles (AGVs)
- Consumer robots
- Industrial robots
- Industry 4.0 applications
Core Capabilities
The platform accelerates common robotics perception functions and provides ready access to:
- Stereo processing — up to 4Kp30 or multiple 1080p30 stereo pairs on the CV2
- Object detection
- Key points extraction and tracking
- Occupancy grid generation
- Visual odometry
- Neural network processing
- OpenCV (Open Source Computer Vision Library) function acceleration
The CV2 chip supports up to 6 direct camera inputs, while the CV25 supports up to 3, enabling multi-camera robotics designs without additional discrete processors.
Jerome Gigot, Senior Director of Marketing at Ambarella, described the platform as combining the company’s advanced imaging capabilities with high-performance CVflow computer vision processing — the goal being smarter and more efficient robots for both consumer and industrial markets.
Software and OS Support
On the software side, the platform supports:
- Linux — for general-purpose robotics applications
- ThreadX RTOS — for products requiring functional safety
- ROS (Robotics Operating System) — for easier algorithm development and visualization
A complete toolkit covering image tuning, neural network porting, and computer vision algorithm development is included. Development kits are available in mono and stereo configurations, with rolling shutter, global shutter, and IR sensor options.
AWS Integration: Optimizing ML Models With Amazon SageMaker Neo
The Problem: Manual Optimization Was a Bottleneck
Before this integration, developers targeting Ambarella CVflow SoCs had to manually optimize machine learning models for the hardware. This was a time-consuming, error-prone step that added friction to the development cycle and slowed time to production.
The Solution: Automated Compilation via SageMaker Neo
Ambarella and AWS collaborated to integrate the Ambarella toolchain directly into the Amazon SageMaker Neo cloud compilation service. The workflow now looks like this:
- Build a model using MXNet, TensorFlow, PyTorch, or XGBoost.
- Train the model in Amazon SageMaker (cloud) or locally.
- Upload the model to your AWS account.
- Use Amazon SageMaker Neo to compile and optimize the model — selecting CV2, CV22, or CV25 as the compilation target.
- Download the compiled model and deploy it to Ambarella-equipped devices.
SageMaker Neo applies a series of compiler optimizations that Ambarella claims can make the resulting model run up to 2× faster on the CVflow neural network accelerator compared to an unoptimized model.
Lightweight Runtime for Edge Deployment
The compiled model runs on the Amazon SageMaker Neo runtime, a purpose-built runtime available in the Ambarella SDK. According to Ambarella, this runtime occupies less than 10% of the disk and memory footprint of full TensorFlow, MXNet, or PyTorch runtimes — a significant advantage when deploying ML inference on resource-constrained connected cameras and edge devices.
Why This Matters for DevOps and MLOps Teams
The Ambarella–AWS integration is a practical example of the broader trend toward cloud-assisted edge ML deployment. For teams building computer vision products on Ambarella hardware, the SageMaker Neo path removes a specialized manual tuning step from the deployment pipeline, replacing it with an automated cloud compilation stage. This fits naturally into CI/CD and MLOps workflows where models are retrained regularly and need to be redeployed quickly to a fleet of edge devices.
The support for multiple frameworks (TensorFlow, PyTorch, MXNet, XGBoost) also means teams are not locked into a single training environment, preserving flexibility earlier in the ML pipeline while standardizing the deployment target.
Availability
As of January 2020, the CVflow robotics platform and related development kits were listed as available. The Amazon SageMaker Neo compilation support for Ambarella CVflow SoCs (CV2, CV22, CV25) was made available through the standard AWS SageMaker service at the same time.