AVSLP: A Multidimensional Evaluation Framework for AI System Performance and Resource Efficiency
4. AVSLP: A Multidimensional Evaluation Framework for AI System Performance and Resource Efficiency This paper introduces AVSLP, a proposed multidimensional evaluation framework for examining artificial intelligence system performance across five dimensions: Actions per Second, Value of Actions, Sustainability, Latency, and Power Consumption. The framework addresses the limitations of evaluating AI systems through isolated performance measurements. High throughput, for example, does not necessarily indicate high-quality outcomes, while reduced power consumption does not automatically establish improved energy efficiency per useful task. AVSLP proposes examining these dimensions together to develop a more complete understanding of system behavior, computational efficiency, and practical performance. It outlines measurement concepts, mathematical representations, evaluation procedures, and reporting recommendations intended to support consistent comparisons between AI architectures and operating strategies. The framework is designed to accommodate different AI systems, workloads, and deployment environments while preserving the distinctions between performance, output value, energy efficiency, latency, and electrical power. AVSLP is presented as a proposed evaluation methodology requiring further testing, refinement, and independent validation.
Authors
- Devin "River" Bigelow
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-13
- DOI
- https://doi.org/10.5281/zenodo.24209291
- Primary Topic
- Parallel Computing and Optimization Techniques
- Type
- preprint