Authors:  V Rajaraman1,∗, Balasubramanian 2 and R. Vaidyeswaran 3

JSA-Vol. 4 (2025),

1 Indian Institute of Technology Kanpur (IIT Kanpur) and the Indian Institute of Science (IISc) Bangalore, India.

2 Indian Institute of Science (IISc) Bangalore and the Indian Institute of Technology Bombay (IIT Bombay), India.

3 Jawaharlal Nehru University (JNU) and the University of Delhi Delhi.

* Correspondence: rajaraman.v37@gmail.com

Received: 10 November 2024; Accepted: 9 June 2025; Published: 12 July 2025.

Abstract: The achievement of exascale performance by systems such as Frontier marks a historic milestone in high-performance computing (HPC). However, this achievement has also exposed a fundamental limitation: performance scaling at exascale is increasingly constrained by energy consumption rather than raw computational capability. Frontier delivers over one exaFLOP of performance while operating near a 20–21 MW power envelope, highlighting the urgent need for energy-aware system design beyond exascale. This paper investigates energy-efficient post-exascale computing through an integrated AI–HPC co-design perspective. We analyze architectural, algorithmic, and runtime-level strategies that jointly optimize performance and energy efficiency on heterogeneous CPU–GPU systems. In particular, we examine mixed-precision computing, AI-assisted scheduling, energy-aware runtimes, and application–hardware co-design as key enablers for sustainable post-exascale systems. By synthesizing insights from recent exascale deployments and AI-driven optimization techniques, this work provides a structured research roadmap for achieving higher scientific throughput under strict power constraints.

Keywords: Exascale computing, Energy efficiency, AI–HPC convergence; Heterogeneous systems, Post-exascale architectures