Tri-Accel: Curvature-Aware Precision-Adaptive and Memory-Elastic Optimization for Efficient GPU Usage
Recommended citation: Sheibanian, M., Shaeri, P., Beigi, A., Woo, R. T., & Keluskar, A. (2025). "Tri-Accel: Curvature-Aware Precision-Adaptive and Memory-Elastic Optimization for Efficient GPU Usage." arXiv:2508.16905.
Published in arXiv preprint; submitted to ScaleOPT @ NeurIPS 2025, 2025
We present Tri-Accel, an optimizer that dynamically adapts precision and memory allocation using curvature-aware updates to maximize GPU efficiency for large-scale training workloads.
