ABSTRACT
Deploying AI at the edge can be challenging. AI algorithms are very compute intensive. In the data centre, multiple large, power-hungry GPUs are often employed. However, edge systems typically have constrained compute capabilities and limited power. Further, many systems need to deal with size, weight, cost, thermal, and other limitations. Successfully deploying AI while meeting these limitations requires a holistic analysis of the system. A Model-Based Cybertronic System Engineering (MBCSE) methodology enables modelling and analysis of complex systems at a high abstraction level. It can be used to analytically find an optimal system architecture and hardware/software partitioning. Meeting the computational requirements may call for the development of bespoke machine learning accelerators. These are complex dedicated compute resources that deliver parallel computation, local data buffers, and some level of programmability. Designing an optimal accelerator architecture can be accomplished with an AI assisted High-Level Synthesis (HLS) process to efficiently explore the design space.
This paper describes a Model Based Cybertronic System Engineering (MBCSE) methodology that can be used to craft a combined hardware/software implementation of an inferencing algorithm, balancing performance, power, cost, and other key design metrics. It begins with an algorithmic analysis, determining areas of significant complexity. This is followed by allocation of functions to physical computation elements, targeting 112both board-level and chip-level placement. During the allocation phase, complex algorithms may be mapped to bespoke accelerators that will be synthesized from the algorithmic description using high-level synthesis. Finally, an analysis of design is performed to ensure that all design metrics are met.
