A STUDY OF GPGPU COMPUTATIONAL DEVELOPMENT FOR EMBEDDED DEVICES

Abstract

  Embedded devices are increasingly demanding high computational power for AI and ML applications. This paper explores the architectural, operating system, and software development challenges associated with integrating General-Purpose computing on Graphics Processing Units (GPGPU) into embedded systems. The paper highlights the diverse approaches adopted by vendors to bring GPUs closer to CPUs. It discusses the challenges of real-time computing deadlines due to the lack of hardware preemption in GPGPUs and explores potential software solutions. The limitations of current operating system support for GPGPUs are addressed, emphasizing the lack of control over execution context compared to CPUs. The role of frameworks like CUDA and OpenCL in facilitating GPGPU programming and integrating with other computing devices (FPGAs, TPUs, DSPs) is explored. The paper advocates for OpenCL as a more widely accepted platform for GPGPU computations, contrasting it with CUDA's hardware specificity. The challenges faced by programmers in choosing frameworks and designing applications for diverse hardware and software environments are discussed. The paper concludes by outlining a focus on three areas for further exploration: architectural development of devices, operating system adaptations, and the evolution of frameworks for GPGPU programming in embedded systems.