Feb 1, 2018
Qualcomm products mentioned within this post are offered by Qualcomm Technologies, Inc. and/or its subsidiaries.
Do you know anyone who doesn’t own a smartphone? By 2020, that could be nearly impossible as 7.8 billion people will be on the planet and a whopping 6 billion smartphones will be in circulation, generating $355 billion in revenue, according to IHS Markit.
As consumers’ passion for smartphones continues to soar, so follows the demand for on-device AI capability. In a new report from CSG Systems International, young adult millennials identified AI as being a major priority for their smartphones, with a majority (53 percent) of respondents noting that they’d spend more for a device equipped with AI features and functionality.
In two years, 6 billion smartphones will be in operation – and many of them just might have AI.
Qualcomm Technologies Inc. and its customers are leading the way in on-device AI, and our Snapdragon mobile platform has hastened the migration of intelligence from the cloud to edge devices in key device categories, including mobile, IoT, and automotive. The Qualcomm Snapdragon 845 Mobile Platform is our third generation mobile AI platform and it’s been optimized to significantly improve your processing speed and boost your mobile experiences related to camera, gaming, XR, voice recognition, and more.
Snapdragon 845 introduces our new Qualcomm Hexagon 685 Vector DSP architecture, plus GPU and CPU optimizations, that together deliver up to three times faster processing of neural networks running on-device compared to the prior generation SoC. And because your information is processed on device and not on the cloud, your experiences will be blazing fast, incredibly connected, and security-rich.
Snapdragon 845 supports multiple frameworks such as Google’s TensorFlow, TensorFlow Lite and Facebook’s Caffe2. Additionally, we’re adding support for the ONNX (Open Neural Network Exchange) interchange format in the Snapdragon Neural Processing Engine (SNPE) SDK to help make it easy for developers to develop neural networks in the framework of their choosing and select the Snapdragon compute core that matches the power and performance profile of the features they deliver on device. In other words, our AI platform is the largest in terms of support for frameworks and partners working to deliver your next AI experience.
SDK helps developers select the optimum compute cores (GPU, CPU, and DSP) to run their apps on Snapdragon.
Supporting platform-wide optimizations
Moving AI inference from the cloud to the mobile device, Snapdragon 845’s new core architectures – our Hexagon DSP, Qualcomm Adreno GPU visual processing subsystem, and Qualcomm Kryo CPU – are working to run AI applications quickly and efficiently. This heterogenous computing approach gives developers expanded choices to support a myriad of neural networks and precisions, high quality apps, and ultimately superior compelling user experiences.
Snapdragon 845 is designed to support numerous neural network frameworks as well as a broad variety of precisions across our three cores. Typically, a neural network model is trained in the cloud at Floating Point 32 (high precision). It could similarly be run on-device on the CPU or GPU, but this is costly and the device pays a penalty in terms of performance and energy. In Snapdragon 845, we’ve introduced optimizations for Floating Point 16 precision on the GPU and optimized support for 8-bit fixed point precision on the CPU, both of which help give developers choices in how to deliver their AI-powered user experiences.
Additionally, our HVX (Hexagon Vector Extensions) is a vector acceleration engine within the DSP subsystem that makes it possible to accelerate neural network execution with 8-bit vector math. In Snapdragon 845, we’ve created a micro architecture that doubles 8-bit performance compared to the prior generation.
Creating superior immersive user experiences
Snapdragon 845 supports a myriad of AI applications like face detection, bokeh effects, and landmark detection for companies like Motorola, OPPO, and Vivo. Vision-based neural networks can be used in place of some traditional camera components. For example, creating a bokeh effect was typically performed with a fixed CV algorithm and two cameras. Snapdragon 845 is engineered to make it easy to utilize a neural network paired with one camera to achieve enhanced performance in terms of latency and accuracy.
Growing the AI developer ecosystem
Our core value is our compute hardware and providing the tools like the NPE SDK to harness that compute. In some instances, we’re developing purpose-built apps like bokeh effects.
To further push the AI envelope and accelerate new and dynamic applications of on-device AI, we’ve assembled an ecosystem of developers that can deliver the use cases, underlying models, and applications that are pre-optimized for Snapdragon 845. Companies like Sensetime, Face++, Thundercomm, Elliptic Labs, and Uncanny Vision are already delivering optimized use cases to our customers and more are on the way.
Pushing the envelope in AI, Snapdragon 845’s third generation mobile AI platform makes your smartphone smarter and energizes your mobile experiences, giving you superior speed, efficiency, and exciting new features.