Sarcouncil Journal of Engineering and Computer Sciences

Abstract: To accommodate the exponentially growing need for artificial intelligence systems, the environment of data centers and cloud infrastructure is experiencing a fundamental change. This flow of technical innovation cuts across several important areas: purpose-built silicon designs to be optimized to only compute AI models, high-bandwidth interconnect technologies to enable parallelization and distribute these models across cluster accelerator networks, breakthroughs in memory technologies to address underlying bandwidth constraints, and advanced parallelism approaches to allow scaling models to massively distributed hardware. Most modern-day AI accelerators employ special-purpose tensor processors and matrix processors that are far more effective at neural network tasks than general-purpose computing, as well as being less power hungry. High-end interconnects, such as NVLink, CXL, and UCIe, buffer out communication overheads between accelerators, and Advances in High Bandwidth Memory enable the astronomical data access requirements in large language models. Such architectural steps are coupled with the increased invasion of environmental awareness, and one of the frameworks was developed to find a compromise between computational power and the sustainability issue. The merge of these domain-specific technologies implies a paradigm change to move beyond the traditional computing architectures in the direction of tailored architectures (which are built specifically to address computational signatures of artificial intelligence workloads).