Compared to Huawei, Apple, and MediaTek, which have different approaches to endpoint learning computing acceleration, Qualcomm, while believing that the use of independent learning computing components to improve efficiency is also important, currently focuses on leveraging existing CPUs, GPUs, and other computing components with software and learning frameworks to achieve equivalent or even superior machine learning computing capabilities. This approach is also less limited by the capabilities of independent computing component versions, allowing more devices based on the Snapdragon computing platform to quickly deploy AI applications.
In an interview with Gary Brotman, Qualcomm's Director of Product Management for Artificial Intelligence and Machine Learning, the company further elaborated on Qualcomm's current approach to on-device machine learning applications and its views on the deployment of AI technology in mobile devices. Brotman also stated that on-device AI applications and deep learning will become a future trend. This will not only enhance device computing efficiency and reduce the limitations of heavy reliance on cloud-based collaborative computing, but will also ensure personal privacy and security on devices, enabling more efficient device performance and lower power consumption.
Qualcomm's current on-device AI technology deployment focuses on three hardware components: the Kryo CPU, Adreno GPU, and Hexagan DSP, each featuring a custom core architecture. These components integrate with software integration into the Snapdragon NPE (Neural Network Engine) SDK, Android NN API, and Hexagan NN API. They also support common learning frameworks such as Facebook's Caffe and Caffe 2, Google's TensorFlow and TensorFlow Lite, and ONNX (Open Neural Networking Exchange), a framework jointly promoted by Microsoft and other vendors. This facilitates on-device learning computing applications and delivers even more efficient performance.
Qualcomm is also collaborating with service providers including SenseTime and Face++. In cloud applications, Qualcomm is collaborating with Tencent (QQ), Baidu (Duer OS, voice recognition), Amazon and Microsoft (ONNX), Google (TensorFlow Lite, object recognition), and Facebook (Caffe 2, augmented reality applications) to expand and implement specific applications of on-device learning computing.
Regarding Qualcomm's own device-side learning computing application model, it is compared with Huawei, Apple, MediaTek and other manufacturers.Adopt a patternObviously different, Gary Brotman said that this does not mean that Qualcomm is not optimistic about the importance of independent learning computing components. It is just based on the development of technology expansion and more cooperation opportunities. He believes that the endpoint learning computing model constructed by the existing Kryo CPU, Adreno GPU and Hexagan DSP in the computing platform, combined with software and learning framework, is more flexible in development. At the same time, it can also present highly compatible performance on a wider range of hardware devices and ensure data interoperability and performance optimization.
Gary Brotman believes that achieving acceleration through independent learning computing components is not a bad idea. However, in the long run, hardware updates may incur additional R&D time and expenses, and compatibility issues must also be considered. Compared to Qualcomm's current design, only adjustments to the software and learning framework are required to achieve similar or even higher learning computing performance using existing hardware resources. Even if the computing platform version is updated, the original artificial intelligence computing service will not experience incompatibility issues due to hardware version factors.
For Qualcomm, helping more partners create more products that connect to AI applications is a key priority. Gary Brotman believes that Qualcomm doesn't discount the advantages of using independent learning computing elements to achieve acceleration, but rather, considering the risks involved, is focusing its development on AI applications powered by software and learning frameworks, thereby embracing more flexible AI deployment. However, Qualcomm also embraces the design model of introducing independent learning computing elements for specific application needs.
The addition of the Snapdragon 700 series is intended to enable more devices to have artificial intelligence computing capabilities.
As for the newly announced integration of artificial intelligence computing capabilities between the Snapdragon 800 series and the 600 series,New Snapdragon 700 series computing platformIn addition to further segmenting product applications, Qualcomm also hopes to enable more application products to introduce artificial intelligence computing performance, especially in the Chinese market, where demand also hopes to connect more application models.
With the Snapdragon 600 series now also incorporating AI applications, does this mean they'll continue to be incorporated into the 400 and 200 series computing platforms in the future? Gary Brotman stated that the entire Snapdragon computing platform architecture is compatible with the Snapdragon NPE and Hexagan NN API. If Google releases Android P (9.0) with native integration of the Android NN API, it will also enable devices on different computing platforms to be driven in AI computing mode.
The most notable difference is that the 400 and 200 series processors do not incorporate the Hexagan DSP design, resulting in a potential performance difference in overall learning computing. However, this ultimately depends on the actual application of the device; not every learning computing method requires instant response efficiency. Furthermore, Qualcomm may also incorporate components such as the Hexagan DSP into the 400 and 200 series computing platforms in the future, thereby enabling different learning computing application models.
In the future, we will launch a benchmarking tool that can fairly evaluate the AI computing capabilities of devices.
In addition, in response to many people's doubts about the effectiveness of different AI technology applications, Gary Brotman revealed that Qualcomm is currently working on creating a testing tool that can be used for fair comparison. This allows partners to use this to judge the efficiency, accuracy, power consumption performance, and support for learning framework models of different AI computing platforms, thereby deciding to adopt individual computing platforms for different application requirements.
Futuremark recently announced plans to release a similar benchmark tool that can evaluate the performance of artificial intelligence applications based on deep learning efficiency. However, finding a suitable benchmark across numerous learning frameworks, acceleration modes, and hardware variations, and establishing a unified scoring standard, is a complex and time-consuming undertaking.
However, Gary Brotman did not specify the expected launch schedule for Qualcomm's evaluation tool for device-side learning computing performance, nor did he provide further details.
