Hexagon is a trademarked name for a series of digital signal processors and neural processing units developed by Qualcomm and included as a coprocessor in its Snapdragon line of systems-on-chip. The development of this coprocessor is part of the move to enhance the multimedia and signaling processing capabilities of the Snapdragon platform. It was first introduced in 2006 under the QDSP6 branding and as a component of Snapdragon S1 and later adopted the Hexagon DSP branding in 2012 with the introduction of Snapdragon S4. Nonetheless, with the latest developments in chip architecture, Hexagon has evolved to become a versatile digital signal processor. Newer Hexagon processors also work as neural processing units and have effectively become artificial intelligence accelerators.
Pros of Qualcomm Hexagon: Advantages and Main Applications
Qualcomm has named its proprietary digital signal processors with the “Hexagon” branding to reflect the efficiency of its architecture. Note that a hexagon is known for being one of the most geometrically efficient shapes because they can be packed together with minimal wasted space and maximum material strength. Hence, in branding its DSPs and NPUs based on this shape, Qualcomm is underscoring their compactness and versatile applications. The following are the advantages and applications of Qualcomm Hexagon DSPs and NPUs:
1. Specialized Processing Capabilities For Heterogeneous Computing
The main advantage of Hexagon is that it specifically excels in specialized processing like digital signal processing and image signal processing. Its inclusion in Snapdragon chips is intended to support heterogeneous computation. Hence, because it processes specific tasks, it offloads workloads from the CPU and GPU and frees them up for other operations.
A digital signal processor handles the real-time processing or manipulation and transformation of digital signals. These digital signals are numerical representations of analog signals sound, video, or sensor data. Processing these signals requires repetitive mathematical operations. An image signal processor handles the capturing and processing of images and videos.
Take note that digital signal processors are used in processing signals in wireless communications, compressing and decompressing audio and video data, and processing audio data. Image signal processors are used in computational photography features like image enhancement, auto-focus and auto-exposure, high dynamic range processing, and depth sensing.
2. Dedicated Artificial Intelligence Accelerator For AI-Related Tasks
Modern versions of Hexagon processors are equipped with Hexagon Vector eXtensions or HVX and the Hexagon Tensor Accelerator or HTA. These are dedicated AI accelerators designed to boost the performance of tasks related to artificial intelligence processing or applications and equip devices with native AI processing capabilities or on-device AI inference.
HVX is focused on vector processing in which data sets are processed in a series of vector operations. These are common in multimedia and AI tasks. HTA is a tensor processing unit that boosts tensor operations or multi-dimensional array operations to handle complex mathematical operations. These are essential in deep learning and neural network workloads.
Nevertheless, because of its built-in AI accelerators, another key advantage of Qualcomm Hexagon is that it works with the CPU and GPU to support features that require computer vision and machine learning. These include face detection, augmented reality, computational photography, natural language processing, and other generative AI applications.
Cons of Qualcomm Hexagon: Disadvantages and Key Limitations
Remember that the inclusion of a Hexagon as a component of a Snapdragon chip enables heterogeneous computing architecture. This allows the chip to exploit the strength of each processor type and improve overall processing performance. However, because Hexagon is a coprocessor, it has limited general-purpose applications. Qualcomm specifically designed this processor to work in tandem with other processors like the CPU and GPU. The following are the disadvantages and limitations of Qualcomm Hexagon DSPs and NPUs:
1. Complex Software Optimization and Limited Software Ecosystem
Not all software or apps can leverage the advantages of Qualcomm Hexagon. Developing an app to utilize the capabilities of this coprocessor requires optimizing code for the vector processing features of the HVC component and tensor operations of the HTA component. Developers need to use the proprietary development tools and libraries from Qualcomm.
It is also important to underscore the fact that there are different brands of systems-on-chips used in Android devices. Each has its own chip architecture. The entire Android hardware ecosystem is considered fragmented, unlike the tight hardware-software integration of Apple. Some developers can only allocate resources for developing for a particular chip platform.
Most third-party developers focus on optimizing their code for CPUs and GPUs that are more universally supported across different chip platforms and devices. Nevertheless, because of these drawbacks, the software ecosystem for Hexagon is limited. Only devices equipped with Snapdragon systems-on-chip can benefit from Software optimized for Hexagon.
2. Limited and System-Dependent Artificial Intelligence Accelerator
The newer versions of Hexagon equipped with Hexagon Vector eXtensions and Hexagon Tensor Accelerator make them dedicated AI accelerators. However, compared with other implementations from competitors, such as Apple and Huawei, the entire processor is not a standalone component for accelerating tasks or workloads related to artificial intelligence.
Apple has a dedicated and independent AI accelerator called Neural Engine. This is found in its A series of chips used in iPhones and certain models of iPads, and M series of chips used in Macs and higher-end models of iPads. Huawei has a dedicated Neural Processing Unit in its Huawei Kirin chips while Google has a Tensor Processing Unit in its Tensor chips.
The implementations of companies like Apple and Google are different from Qualcomm. Their artificial intelligence accelerators are not system-dependent. The approach of Qualcomm centers on using Hexagon alongside another AI accelerator called Sensing Hub, the Kryo CPU cores, and Adreno GPU. These four processors form the Qualcomm AI Engine.