• Technical Conference:  30 March – 03 April 2025
  • Exhibition: 01 – 03 April 2025
  • Moscone Center, San Francisco, California, USA

Advanced Packaging and Integrated Optics for Scale-Up AI interconnects

Moscone Center

Artificial Intelligence (AI) has been driving development of state-of-art semiconductor technology within compute, memory, storage and networking domains over the last several years. Most recently however, Generative AI (Gen AI) applications, such as ChatGPT and Sora, have provided a massive impetus to the demands for computational power. Specifically, the scaling up of GPU (Graphics Processing Unit) clusters faces challenges in bandwidth density and reach. Electrical interconnects and SERDES technology are starting to run out of steam as GPU and switch die-edge bandwidth densities surge past 1 Tbits per second/mm (Tbps/mm). The number of GPUs per cluster is also growing, which begs for cluster interconnect to span multiple racks at reasonable power levels. The large impact of distance to the energy and latency of electrical interconnects is a significant impediment to scaling up.  

This symposium focuses on the widespread efforts across industry and academia to bring the optical transceiver technology, which is now ubiquitous in the datacenters, closer to the compute chips. These efforts are either in the form of Co-Packaged Optics (CPO) or Optical IO, which bring the Photonics even closer to the chip using  Advanced 2.5-3D Packaging technologies.

The symposium will be split into two sessions.

  • The first session will be kicked off with an introduction to the existing solutions and their limitations vis-à-vis the diverse application scenarios (training vs. inference, B2B vs. B2C, etc.). This will motivate the required metrics to compare different optical architectures (wide and slow, vs. fast and narrow) and optical platforms (Single vs. multi-mode, VCSELs vs. Silicon Photonics, 2D vs. 2.5D vs. 3D packaging platforms).
  • The second session will dig into the pros and cons of the different optical architecture and platform choices. It will delve into the choices of external vs. internal lasers, ring vs. Mach-Zehnder modulators, optical packaging options and all associated ecosystem developments from foundry to OSATs. It will further address considerations for system integration into GPU servers, such as associated thermal management tradeoffs, and reliability and serviceability requirements.

The format of the symposium will encompass sufficient time for open discussions as the industry navigates this multi-dimensional and multi-physics effort.

Karl Bois, NVIDIA, USA
Daryll Childers, US Conec Ltd, USA
Julie Eng, Coherent, USA
Shan Gao, TSMC COUPE, Taiwan
Sylvie Menezo, SCINTIL Photonics, France
Peter O’Brien, Tyndall, Ireland
Noam Ophir, Jabil, USA
Vivek Raghuraman, Mixx Technologies Inc, USA
Anand Ramaswamy, Broadcom, USA
Hesham Taha, Teramount Ltd, Israel