Year: 2026
Type: article
Abstract: This article investigates the methods for increasing the efficiency of the inference stage in modern artificial intelligence systems, particularly in Large Language Models and deep neural networks. The paper provides a comparative analysis of the hardware and software hierarchy of two distinct Domain-Specific Architectures designed to overcome the limitations of general-purpose graphics processors: Google's Tensor Processing Unit and Groq's Language Processing Unit. The impact of Systolic Arrays and Software-Defined Deterministic Scheduling mechanisms on inference speed, latency, and energy efficiency is substantiated using mathematical models. Simulation and benchmark results indicate that the Language processing unit architecture reduces latency in sequential token generation by up to 8–10 times compared to graphics processors platforms, while the Tensor processing unit delivers higher throughput and energy savings (up to 26.8%) in parallel matrix computations.
Institution: Bukhara State University
Language: English
Subfield: Hardware and Architecture
Field: Computer Science
Domain: Physical Sciences
Open Access status: diamond
OpenAlex ID: W7220311881