This blog post was originally published at Ambarella’s website. It is reprinted here with the permission of Ambarella.
Applied machine learning and artificial intelligence are quickly becoming the new horizon of robotics applications, rapidly expanding the possibilities of automated solutions. At Ambarella, our SoCs are at the intersection of high performance, low power, and low cost of ownership. Powered by proprietary CVflow® AI hardware accelerators, our SoCs enable high-performance neural networks capable of making a developer’s vision come true—enabling the creation of the next generation of robotics applications at scale.
Compared to existing GPUs and FPGAs, Ambarella SoCs offer higher computer vision performance per watt, easy integration with a multitude of 3D depth sensors, a lower thermal envelope, and easier migration— supported by a unified SDK approach—providing a high degree of flexibility in choosing the right SoC for each application.
Ambarella Robotics: What Sets Us Apart
At Ambarella, video and compression are in our DNA. We understand that quality in is quality out. Whether it be CMOS image sensors, or other sensors such as LIDAR, RADAR, and IR, Ambarella’s proprietary ISP consistently delivers industry-leading image processing quality. Regardless of the sensor modality, whether it be passive stereo, active stereo, monocular depth, time of flight (ToF), or structured light, our ISP ensures high image quality across a wide range of lighting conditions.
Neural Network Accelerator – CVflow
Our unified software infrastructure enables rapid development of robotics perception applications. Ambarella’s CVflow technology provides a dedicated computer vision accelerator capable of running multiple neural network models simultaneously and efficiently, accommodating a wide range of approaches to computer vision, and giving developers a high degree of flexibility. CVflow supports a variety of popular open-source frameworks such as PyTorch, Tensor Flow, Caffe2, and ONNX. Our CNNGen toolchain makes it easy to port existing neural networks onto our EVK, enabling image analysis that consistently exceeds expectations.
While our SoC is performing complex computations, our multicore Arm® CPU makes it possible to supervise the overall chipset while maintaining the ability to run additional algorithms such as path planning, navigation, and SLAM, removing the need for additional chipsets.
Our SoC is capable of running high-compute, complex algorithms while keeping power consumption low, providing best-in-class thermal shielding to the system design. This reduction in battery power draw and heat generation allows us to pass on cost-savings to consumers by removing the need for large heat sinks or extensive use of thermal pastes, making it an ideal SoC for processing at the edge.
At Ambarella, we take security extremely seriously. Beginning at the chip development stage, security features such as secure ROM, secure boot with randomly generated keys stored on the chip’s OTP area, content encryption, DRAM scrambling, and model protection are implemented at both chipset and SDK levels.
Scalable Silicon Solution
It’s time to scale. With our proven track record for delivering smaller-process geometry computer vision (CV) chips, scaling has never been easier. Compared to the high cost of ownership for GPU or FPGA platforms, our higher yield output offers significant cost reductions for our customers. Altogether, our CV chips’ advantages in image quality, computer vision, chip functionality, power consumption, and security features enable customers to develop best-selling consumer and professional robots.
Chipsets Targeted for Robotics
- CV2 / CV2FS / CV5x: Industrial robotics application (AGVs / AMRs) and high-end vacuum cleaners
- CV22x: Machine vision applications
- CV25x: Consumer robotics applications
- CV28M : Low-power, lowest-cost SoC
Below is an animated overview of our approach to consumer robotics.
Global Product Marketing – Robotics, Ambarella