Abstract

The present disclosure provides a method (300) for adaptive semantic memory paging in artificial intelligence (AI) workloads. The method (300) includes receiving an AI workload comprising a plurality of memory pages, monitoring semantic execution signals associated with the AI workload, computing semantic relevance scores for the memory pages based on the semantic execution signals, and prioritizing the memory pages according to the semantic relevance scores. The method (300) further includes determining memory-tier assignments for the memory pages, placing higher-priority memory pages in higher-performance memory tiers, identifying lower-priority memory pages for migration or eviction, detecting memory pressure and changes in model behavior, updating the semantic relevance scores responsive to the detected changes, and reallocating the memory pages across memory tiers based on the updated semantic relevance scores. The disclosed method enables semantic-aware memory management for AI workloads, thereby improving memory utilization, reducing page faults and latency, and supporting efficient execution of large-scale AI models in memory-constrained computing environments.

Creative Commons License

Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.

Download

Share

COinS