Year: 2025
Type: article
Abstract: This article examines the optimization strategies for inference workloads in real-time systems across various performance-critical applications. As artificial intelligence becomes increasingly embedded in time-sensitive domains, the need for efficient execution of inference tasks under strict latency constraints has become paramount. Unlike training processes prioritizing accuracy regardless of computational cost, inference workloads must balance precision with performance constraints, particularly in resource-limited environments. The article explores three key optimization approaches: reduced precision computing techniques that preserve accuracy while decreasing computational demands, resource allocation and workload management strategies that adapt to fluctuating conditions, and specialized cache architectures that minimize memory access latencies for tensor operations. Through case studies in autonomous vehicles, industrial automation, and financial transaction monitoring, the article demonstrates how these optimizations enable mission-critical AI systems to meet stringent real-time requirements. Additionally, emerging directions, including hardware-software co-design, neural architecture search for efficiency, and sparse computation, are explored as promising frontiers for future optimization efforts
Author: Deepika Bhatia
Institution: San Jose State University
Language: English
Topic: Real-Time Systems Scheduling
Subfield: Hardware and Architecture
Field: Computer Science
Domain: Physical Sciences
Open Access status: diamond
OpenAlex ID: W4410072872