Analog AI Accelerators for Transformer-based Language Models: Hardware, Workload, and Power Performance
Transformer-based Large Language Models (LLMs) demand large weight capacity, efficient computing, and high throughput access to large amount of dynamic memory. These challenges present great opportunities for algorithmic and hardware innovations, including Analog AI accelerators. In this paper, we describe recent progress on Phase Change Memory-based hardware and architectural designs to address the challenges for LLM inference.
Authors
- Kohji Hosokawa (ORCID: https://orcid.org/0009-0009-8086-8144)
- Geoffrey W. Burr (ORCID: https://orcid.org/0000-0001-5717-2549)
- Charles Mackin (ORCID: https://orcid.org/0000-0001-8413-5583)
- Atsuya Okazaki (ORCID: https://orcid.org/0000-0002-5275-5224)
- Abu Sebastian (ORCID: https://orcid.org/0000-0001-5603-5243)
- Hadjer Benmeziane (ORCID: https://orcid.org/0000-0002-5259-0749)
- Jose Luquin (ORCID: https://orcid.org/0009-0005-9539-1386)
- Manuel Le Gallo (ORCID: https://orcid.org/0000-0003-1600-6151)
- Kaoutar El Maghraoui (ORCID: https://orcid.org/0000-0002-1967-8749)
- S. Jain
- Yasuteru Kohda
- Julian Büchel (ORCID: https://orcid.org/0000-0001-9495-7150)
- Vijay Narayanan (ORCID: https://orcid.org/0009-0008-8433-963X)
- Pritish Narayanan (ORCID: https://orcid.org/0000-0002-3176-0059)
- Masatoshi Ishii (ORCID: https://orcid.org/0000-0003-0794-7232)
- Alexander Friz
- Irem Boybat (ORCID: https://orcid.org/0000-0002-4255-8622)
- Hsinyu Tsai (ORCID: https://orcid.org/0000-0002-3971-097X)
- William A. Simon (ORCID: https://orcid.org/0000-0002-1832-4690)
- Athanasios Vasilopoulos
- A. Chen (ORCID: https://orcid.org/0009-0005-3814-7264)
Institutions
- IBM Research - Almaden (US)
- IBM Research - Zurich (CH)
- IBM Research - Tokyo (JP)
Publication Details
- Published
- 2025-05-18
- DOI
- https://doi.org/10.1109/imw61990.2025.11026974
- Primary Topic
- Ferroelectric and Negative Capacitance Devices
- Type
- article
- Field-Weighted Citation Impact
- 0.00