Heterogeneous Computing in the Edge
communications traffic , and when certain events are detected the much more powerful companion ARM core wakes up to perform advanced computations . In 2020 , NVIDIA purchased ARM Holdings .
Rounding out the CISC / RISC CPU category is the relatively recent RISC-V architecture 16 , introduced in 2015 . Unlike the X86 and ARM architectures , which are owned and licensed by commercial companies , RISC-V is an entirely open source implementation . This means that manufactures of a System on Chip ( SoC ) can freely import , adapt and modify RISC-V without concern for royalties or intellectual property .
The ecosystem is rapidly developing surrounding RISC-V , with dozens of manufacturers of core intellectual property , SoCs , development boards , and systems . RISC-V may have advantages over X86 and ARM for future edge computing systems by virtue of its highly flexible architecture , expanding supply chain , and potentially lower cost of ownership commercial model .
There are a number of additional CISC / RISC CPU architectures competing for relevance in the edge computing market . For the sake of brevity , I didn ’ t include an exhaustive list here .
A common attribute of all the CISC / RISC processor architectures is their CPU cores are optimized for single thread performance , and they generally do not perform as efficiently on algorithms that can use massive parallelism , or would benefit from specialized data path designs . CISC / RISC CPUs are the obvious choice for control planes , configuration engines , log file maintenance , management processes , and similar applications , and since these are essential to most systems , at least a modest CISC / RISC supervisory subsystem is helpful in most edge computing applications .
Larger multicore CISC / RISC CPUs can be applied to some parallel execution environments , but in general they start to lose efficiency for applications with more than 50-100 parallel execution threads . Beyond that , the addition of some other type of processing resource supporting more massive parallelism or specialized data paths makes sense , and that is the opportunity for heterogenous processors in the edge . That opportunity is the focus of the remainder of this paper .
GPUs
Graphics Processing Units ( GPUs ) are specialized processor architectures that were originally designed to accelerate graphical computing operations like geometric transformation , windowing , clipping , texture mapping , ray tracing , shading , and rendering . They achieve ordersof-magnitude performance improvement over CISC / RISC CPUs for the specialized graphical
16
IIC Journal of Innovation - 49 -