Mar 31, 2021
Qualcomm products mentioned within this post are offered by Qualcomm Technologies, Inc. and/or its subsidiaries.
Are you still running your artificial intelligence workloads in the cloud? That may make sense for training your models, but if your applications depend on techniques like person detection and pose estimation to name a few, then it’s time you looked into on-device AI.
Suppose you have a sports or a fitness application that makes decisions based on user interaction through a camera in a gym. Do you really want to send the camera feed up to the cloud for inference? Think about the latency your app will have to fight, running bits back and forth across the network. Then think about your users’ privacy, as their likeness and the personal data they’ve entrusted to you are getting processed somewhere in the cloud.
With models running on devices at the network edge, you can make predictions and take action where the data resides. That’s faster, more secure and more reliable than running models in the cloud.
Qualcomm Vision Intelligence platform featuring the Qualcomm QCS610 processor, engineered for on-device AI
We’ve made the QCS610 system on chip (SoC) with your application in mind, by combining high performance and low power consumption in a single package.
Designed for the internet of Things (IoT), our QCS610 has the computing power to run AI models in devices like IP cameras, dash cams, 360/VR cameras and wearables. In fact, with the heterogeneous-computing trio of the Octa-core Qualcomm Kryo CPU, Qualcomm Adreno GPU, and Qualcomm Hexagon DSP, you can run multiple AI models simultaneously. This video shows you what that looks like:
As you can see, we have one QCS610-powered device running three models in parallel:
- MobileNetV2 SSD for detection — to indicate that a person is in the field of vision
- DeepLab v3 for image segmentation — to separate the object in the foreground of the image from the background
- PoseNet for pose estimation — to track physical movements in the field of vision
We’ve built into our QCS610 a series of enhanced AI intrinsics in the Hexagon DSP for a 50-percent boost in throughput compared to previous models. That boost helps enable high-quality image processing and video while still taking it easy on power consumption.
Besides its prowess with AI, the QCS610 handles 4K Ultra HD video at 30fps. Its video capture and playback capabilities include advanced noise reduction and low-light performance. It’s equipped with dual image signal processors (ISP), hardware-based security, a video processing engine and audio codecs for multimedia and voice control.
Multiple software stacks for developers
As a developer, you’re as keen for software as you are for hardware, so we’ve lined up tools you can use to get the most out of the QCS610.
- Qualcomm Artificial Intelligence (AI) Engine: The engine combines dedicated hardware and software designed to accelerate the execution of machine learning inference workloads on the device.
- Qualcomm Neural Processing SDK for AI: Our software-accelerated runtime for the execution of deep neural networks lets you program the Qualcomm AI Engine. Use the SDK to develop custom AI models based on industry-standard frameworks such as TensorFlow, TensorFlow Lite, ONNX, and Caffe2 then determine the optimal mix of cores (Kryo CPU, Adreno GPU, Hexagon DSP) for running your workloads. The SDK comes with an easy-to-use reference guide and learning resources.
- AI Model Efficiency Toolkit (AIMET): Getting accuracy from floating-point models in the integer world of embedded applications takes work. Qualcomm Innovation Center, Inc. has released AIMET as open source with a model zoo containing more than a dozen INT8 models, giving you the high accuracy of floating-point math with 8-bit integer efficiency.
- Qualcomm Connected Camera SDK: This audio-visual framework has APIs for encoding multiple channels of AV from the camera, with pre- and post-processing.
The QCS610 SoC offers enterprise-grade AI, video and connectivity (WI-FI 802.11ac, Bluetooth 5.0, Ethernet/RGMII) with low power consumption and solid developer tools. It’s ideal for edge applications like industrial IoT, smart AI enterprise security, home IP Cameras, dash cams, body cams, smart displays and videoconferencing.
Visit the QCS610 product page for specs, documentation, reference designs and development kits that will show you how you can put the QCS610 to work in your products and applications.