| Περίληψη: |
The rapid growth of large-scale artificial intelligence (AI) models has placed memory management at the centre of both hardware and software design challenges. The present study investigates and introduces the Memory Capacity Estimation Model (MCEM): a parametric, device-agnostic, and precision-aware analytical framework for predicting peak memory utilisation across central processing units, graphics processing units, and neural processing units. MCEM formalises a dual-context memory taxonomy that distinguishes training (bandwidth-bound, writedominated) from inference (latency-sensitive, key-value cache-dominated) based on dataflow characteristics, temporal write/read ratios, and reuse distance. Validated against open benchmarks from MLPerf, Hugging Face, and advanced micro devices Instinct MI300X hardware reports, MCEM achieves estimation errors of 0.73%-2.7% across five diverse model-device configurations. The framework is shown to reduce memory overprovisioning by 30%-40% relative to naive sizing, with implications for sustainable, cost-effective AI deployment at cloud, high-performance computing, and edge scales. [ABSTRACT FROM AUTHOR] |