TR2024-105

Hardware-Efficient Quantization for Green Custom Foundation Models

    •  Koike-Akino, T., Meng Chang, , Cevher, V., De Micheli, G., "Hardware-Efficient Quantization for Green Custom Foundation Models", International Conference on Machine Learning (ICML), July 2024.
      BibTeX TR2024-105 PDF
      • @inproceedings{Koike-Akino2024jul2,
      • author = {Koike-Akino, Toshiaki and Meng Chang and Cevher, Volkan and De Micheli, Giovanni}},
      • title = {Hardware-Efficient Quantization for Green Custom Foundation Models},
      • booktitle = {International Conference on Machine Learning (ICML)},
      • year = 2024,
      • month = jul,
      • url = {https://www.merl.com/publications/TR2024-105}
      • }
  • MERL Contact:
  • Research Areas:

    Artificial Intelligence, Machine Learning

Abstract:

We propose a new hardware-efficient quantization (HEQ) for low-power full-custom foundation models. The HEQ jointly optimizes multiplier hardware and weight quantization to minimize the total power consumption. Exploiting power profile of custom multipliers, our method achieves a significant power reduction up to 20 folds.