Heterogeneous Computing in the Edge
One response to this challenge is Google ’ s “ Coral ” series of edge TPU products . Their Dev Board Mini has a throughput of 4TOPS , uses 2 watts of power , has 2GB of DDR3 RAM , has a physical size of 64mm x 48mm , and costs $ 100 in small quantities 27 . It uses a simplified programming environment called TensorFlow Lite .
FPGAs
Field Programmable Gate Arrays ( FPGAs ) are an important processing technology for edge computing . FPGAs consist of a large array of programmable logic blocks interconnected by an array of programmable data paths to implement custom hardware that is ideally matched to specific computational problems . The programmable logic blocks consist of a look-up table that implements logic functions , a storage element for intermediate results , and connections to the chip ’ s programmable interconnect network . Many modern FPGAs expand the programmable logic blocks to also include multipliers , memory blocks , specialized I / O structures , and RISC processors . The programmable logic blocks and interconnect are initialized by a configuration file that is shifted in from external storage and stored in configuration RAM within the FPGA . Most FPGAs can be reconfigured dynamically , so it is possible to change the hardware data paths on the fly as different applications may require .
Two companies dominate the FPGA world , and both have been acquired by semiconductor giants . Altera was acquired by Intel in 2015 28 , and AMD recently announced their acquisition of Xilinx 29 . The combination of a large processor company and a leading FPGA company creates a powerful combination in heterogenous computing . In addition to Altera and Xilinx , there are a number of smaller players in the FGPA space , including Lattice , Microchip , and QuickLogic .
FPGAs are programmed using a hardware design process and high level HW design language . This means the design skills needed to effectively produce the configuration files is closer to a hardware designer than a software programmer . However , the design automation systems for FPGAs continue to mature , making FPGA programming accessible to audiences with less HW experience 30 .
FPGAs are applicable to many different workloads at the edge . Since the FPGA designs are highly optimized for specific applications , their performance and power can be well suited to the constraints of edge computing . Some examples include software defined radio subsystems , high
27
28
29
30
- 52 - June 2021