Heterogeneous Computing in the Edge
performance computing , instrumentation , medical , and financial algorithms . FPGA manufacturers have been producing parts optimized for edge computing , for example the Xilinx Kria product portfolio 31 .
DSPs
Digital Signal Processors ( DSPs ) are specialized computation elements optimized for signal processing workloads . They tend to have parallel arrays of multiply-accumulate engines that are valuable in implementing functions like digital filtering , convolution , FFTs , compression , waveform synthesis , and audio algorithms . Leading processor companies in the DSP space include TI 32 , Analog Devices , Microchip , STMicroelectronics and NXP .
The programming environment for DSPs is often a variant of the environments used in CISC / RISC CPUs . However , knowledge of the specific data paths in the target DSP chip is often required to produce high performing code .
DSPs have some utility at the edge , especially in radio signal and audio processing . However , the algorithms that are traditionally implemented on DSPs are moving to GPUs and FPGAs for many edge implementations , where they can be run at greater efficiency by most measures .
ASICs / ASSPs / SoCs
Rounding out the heterogenous computing at the edge technology universe are Application Specific integrated Circuits ( ASICs ), Application Specific Standard Products ( ASSPs ) and Systems on Chips ( SoCs ). These represent full custom hardware implementations of specific computational capabilities . This is similar to the full custom data paths found in FPGAs , but they are not reconfigurable at the gate level . This makes them faster , smaller , and lower power , but their lack of programmability limits their versatility in diverse edge applications .
ASICs are made for a specific application , typically by the company who is building the system or box-level product . They represent the most optimized , best possible throughput and cost profile for the specific function they are designed for . However , ASICs have serious drawbacks , because it often takes over a year to design one , more than $ 10M to create the masks needed to fabricate the chip , and they are impossible to modify once built . These concerns limit the applicability of ASICs in networks , especially the highly dynamic environments found at the edge .
ASSPs are ASICs designed by semiconductor houses for a certain set of applications , and sold to all of their customers . This eliminates many of the risks associated with full-custom ASICs , but
31
32
IIC Journal of Innovation - 53 -