Regarding the addition of an NPU-like accelerated computing design to the Snapdragon 855 processor's DSP structure starting last year, which is significantly different from the previous AI collaborative computing model composed of a CPU, GPU, and DSP, Qualcomm Senior Director of R&D and Head of the AI Research Program in the Enterprise R&D Department, Hou Jilei, stated that while in the past, emphasis was placed on the collaborative operation of various components within the processor to achieve the effectiveness of AI computing, due to factors such as the continued significant increase in transmitted data and changes in computing requirements, adjustments will also be made to the product architecture design.
Invested in artificial intelligence computing research very early
Qualcomm has been investing in AI-related application research since 2007, accumulating 12 years of research experience. In processor design, it developed the Snapdragon AI Engine (Snapdragon AIE), an artificial intelligence computing engine that combines various computing components including the CPU, GPU, and DSP, and accelerates computing within the Android NN framework.
The same computing architecture design can also be applied to processor products of different levels, making Qualcomm's various processorsBoth can use the same operation mode, allowing app service developers to create content that can be applied to different processor computing environments through a single design, thereby allowing artificial intelligence computing applications to be widely used in various devices.
Adjust computing architecture according to different needs
Driven by the development of 5G network technology, the amount of data processed by devices continues to grow. At the same time, the demand for computing efficiency has also changed, and adjustments must be made to the computing architecture. Therefore, the DSP structure of the Snapdragon 855 processor launched at the end of last year has been added with a similar NPU acceleration computing design, which is actually to achieve better computing efficiency.
However, as Qualcomm Technologies' senior vice president of product departmentKeith KressinSaid that it would notSnapdragon 855The change in DSP structure is considered to be an NPU accelerated computing design. In fact, it still maintains the use of CPU, GPU, DSP and other components, combined with different frameworks to achieve various artificial intelligence computing results. It is believed that even if the DSP structure design changes, it will not affect the existing computing form.
Unlike Huawei and Apple, which integrate dedicated computing acceleration components into their processors, Qualcomm believes that flexibly adjusting the resources of each component according to varying computing needs maximizes efficiency, rather than being limited to a specific computing mode. This design allows devices to adapt to the latest computing technologies at any time, especially as artificial intelligence computing technology continues to evolve.
The development of IoT applications may drive decentralized cluster computing
Qualcomm's current AI computing deployment will shift from a cloud-centric computing model to one where preliminary computing will be completed on the device side in specific situations, and then coordinated with the cloud for computing. As a result, the proportion of computing applications on the device side will continue to increase, ensuring that mobile phone processor performance is sufficient to meet the needs of on-device AI computing.
As for the future computing model, Hou Jilei explained that the current view is that collaborative computing between the cloud and terminal devicesWill be accelerated by 5G network connectionAt the same time, it is expected that the data transmission speed and overall throughput will increase, thus changing the existing computing architecture design. For example, the current collaborative computing model connecting the cloud to the terminal may be changed in the future.Decentralized form, allowing devices to establish connections with each other and even form large computing clusters in the form of arrays.
However, the connection operation between devices still involves data privacy and whether the user is willing to allow personal devices to establish connections with other devices. Therefore, although such an operation model can be achieved technically, it is difficult to implement in practice. Perhaps in the future it can be accomplished through technologies such as blockchain.
