---
_id: '67466'
abstract:
- lang: eng
  text: 'This paper presents a holistic, cross-stack vision for efficient and scalable
    generative AI, motivated by the view that efficiency is most readily improved
    when hardware, compiler, and algorithmic techniques are pursued together rather
    than in isolation. We connect four complementary perspectives: (i) foundational
    hardware innovations, including heterogeneous 2.5D/3D integration and in-memory
    computing, that reduce the compute, memory, and energy footprint of billion- parameter
    models; (ii) extensible compiler infrastructures, built around MLIR, that bridge
    emerging in-memory hardware to real applications and automate design-space exploration;
    (iii) cross-layer and algorithmic techniques –quantization, sparsification, and
    dataflow optimization– that compress and accelerate large language models from
    the cloud to the edge; and (iv) token reduction methods, such as token merging
    and clustering, that attack the quadratic cost of the transformer attention mechanism
    at the model level. By integrating hardware, compiler, and algorithm viewpoints
    into a single narrative, we outline how co-designed solutions across the stack
    can compound to enable affordable, scalable, and democratized access to generative
    intelligence.'
author:
- first_name: Weitian
  full_name: Wang, Weitian
  last_name: Wang
- first_name: Dhingra
  full_name: Pratyush, Dhingra
  last_name: Pratyush
- first_name: Asif Ali
  full_name: Khan, Asif Ali
  last_name: Khan
- first_name: Arkapravo
  full_name: Ghosh, Arkapravo
  last_name: Ghosh
- first_name: Partha Pratim
  full_name: Pande, Partha Pratim
  last_name: Pande
- first_name: Jeronimo
  full_name: Castrillon, Jeronimo
  last_name: Castrillon
- first_name: Priyadarshini
  full_name: Panda, Priyadarshini
  last_name: Panda
- first_name: Cecilia De La
  full_name: Parra, Cecilia De La
  last_name: Parra
- first_name: Shubham
  full_name: Rai, Shubham
  last_name: Rai
citation:
  ama: 'Wang W, Pratyush D, Khan AA, et al. A Cross-Stack Approach to Efficient and
    Scalable Generative AI (Special Session Paper). In: <i>Proceedings of the 2026
    International Conference on Compilers, Architecture, and Synthesis of Embedded
    Systems (CASES)</i>. CASES ’26 Companion. ; 2026.'
  apa: Wang, W., Pratyush, D., Khan, A. A., Ghosh, A., Pande, P. P., Castrillon, J.,
    Panda, P., Parra, C. D. L., &#38; Rai, S. (2026). A Cross-Stack Approach to Efficient
    and Scalable Generative AI (Special Session Paper). <i>Proceedings of the 2026
    International Conference on Compilers, Architecture, and Synthesis of Embedded
    Systems (CASES)</i>.
  bibtex: '@inproceedings{Wang_Pratyush_Khan_Ghosh_Pande_Castrillon_Panda_Parra_Rai_2026,
    series={CASES ’26 Companion}, title={A Cross-Stack Approach to Efficient and Scalable
    Generative AI (Special Session Paper)}, booktitle={Proceedings of the 2026 International
    Conference on Compilers, Architecture, and Synthesis of Embedded Systems (CASES)},
    author={Wang, Weitian and Pratyush, Dhingra and Khan, Asif Ali and Ghosh, Arkapravo
    and Pande, Partha Pratim and Castrillon, Jeronimo and Panda, Priyadarshini and
    Parra, Cecilia De La and Rai, Shubham}, year={2026}, collection={CASES ’26 Companion}
    }'
  chicago: Wang, Weitian, Dhingra Pratyush, Asif Ali Khan, Arkapravo Ghosh, Partha
    Pratim Pande, Jeronimo Castrillon, Priyadarshini Panda, Cecilia De La Parra, and
    Shubham Rai. “A Cross-Stack Approach to Efficient and Scalable Generative AI (Special
    Session Paper).” In <i>Proceedings of the 2026 International Conference on Compilers,
    Architecture, and Synthesis of Embedded Systems (CASES)</i>. CASES ’26 Companion,
    2026.
  ieee: W. Wang <i>et al.</i>, “A Cross-Stack Approach to Efficient and Scalable Generative
    AI (Special Session Paper),” 2026.
  mla: Wang, Weitian, et al. “A Cross-Stack Approach to Efficient and Scalable Generative
    AI (Special Session Paper).” <i>Proceedings of the 2026 International Conference
    on Compilers, Architecture, and Synthesis of Embedded Systems (CASES)</i>, 2026.
  short: 'W. Wang, D. Pratyush, A.A. Khan, A. Ghosh, P.P. Pande, J. Castrillon, P.
    Panda, C.D.L. Parra, S. Rai, in: Proceedings of the 2026 International Conference
    on Compilers, Architecture, and Synthesis of Embedded Systems (CASES), 2026.'
date_created: 2026-10-09T08:46:07Z
date_updated: 2026-10-21T09:14:56Z
publication: Proceedings of the 2026 International Conference on Compilers, Architecture,
  and Synthesis of Embedded Systems (CASES)
series_title: CASES ’26 Companion
status: public
title: A Cross-Stack Approach to Efficient and Scalable Generative AI (Special Session
  Paper)
type: conference
user_id: '131313'
year: '2026'
...