EUVIMEDEUVIMED entdecken
Open Access · CC BY-NC-ND 4.0

PHONON: Silicon-Photonic Network-on-Interposer for Energy-Efficient and High Performance LLM Inferencing

Gaurav Narang, Ashish Reddy Bommana, Krishnendu Chakrabarty

ACM Transactions on Design Automation of Electronic Systems · 31.03.2027 · 32 (2) · 1-29

DOI: 10.1145/3831707 ↗

Kurz gefasst

Fachartikel in der Zeitschrift ACM Transactions on Design Automation of Electronic Systems, online erschienen am 13.11.2026. Verfasst von Gaurav Narang und 2 weiteren Autor:innen. Die Arbeit wurde bisher 1-mal zitiert (Crossref) und nennt 76 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY-NC-ND 4.0).

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Large Language Models (LLMs) exhibit extreme compute and memory demands, necessitating large-scale multi-chiplet systems to achieve high-performance and scalability. In such systems, frequent off-chip memory accesses for model parameters and intermediate activation exchanges induce substantial data movement between compute and storage chiplets, making inter-chiplet communication a critical performance and energy bottleneck. Silicon photonic (SiPh) Networks-on-Interposers (NoIs) offer high bandwidth and low latency communication, making them an attractive choice for the inter-chiplet communication backbone. However, conventional SiPh NoI designs require long-range multi-hop data exchanges, resulting in degraded performance and energy efficiency, and poor scalability. In this work, we present PHONON, a novel Silicon- Pho tonic N etwork- o n-I n terposer architecture, that jointly optimizes compute–compute and compute–storage inter-chiplet communication. PHONON employs structured waveguide placement, cluster-level connectivity to storage chiplets, and optimized mapping of compute kernels to chiplets. PHONON optimizes waveguide crossings and reduces waveguide propagation losses, insertion losses, and memory-access contention. Simulation results show that PHONON reduces NoI latency and energy consumption by up to 9.4× and 28.4×, respectively, compared to state-of-the-art SiPh NoI architectures, enabling scalable, high-performance, and energy-efficient LLM acceleration. PHONON reduces the LLM-serving metrics of time-to-first-token (TTFT) by up to 1.48×, and improves tokens/sec and tokens/Joule by up to 4.9× and 28×, respectively, compared to SiPh NoI counterparts.

Zitieren

APA

Narang, G., Bommana, A. R., & Chakrabarty, K. (2027). PHONON: Silicon-Photonic Network-on-Interposer for Energy-Efficient and High Performance LLM Inferencing. ACM Transactions on Design Automation of Electronic Systems, 32(2), 1-29. https://doi.org/10.1145/3831707

Vancouver

Narang G, Bommana AR, Chakrabarty K. PHONON: Silicon-Photonic Network-on-Interposer for Energy-Efficient and High Performance LLM Inferencing. ACM Transactions on Design Automation of Electronic Systems. 2027;32(2):1-29. doi:10.1145/3831707

BibTeX
@article{Narang2027,
  author = {Narang, Gaurav and Bommana, Ashish Reddy and Chakrabarty, Krishnendu},
  title = {PHONON: Silicon-Photonic Network-on-Interposer for Energy-Efficient and High Performance LLM Inferencing},
  journal = {ACM Transactions on Design Automation of Electronic Systems},
  year = {2027},
  volume = {32},
  number = {2},
  pages = {1--29},
  doi = {10.1145/3831707}
}
Kostenlos starten →