Sunday, 06 March
13:00 – 15:30
Organizers:
Nicola Calabretta, Eindhoven Univ. of Technology, Netherlands
Wenhua Lin, Intel Corp., USA
Michela Svaluto Moreolo, Ctr Tecnològic de Telecom de Catalunya, Spain
Description:
Hyperconnectivity is driving a radical change in our society and posing several technological challenges, fueling changes in communications because of the complexity, diversity and integration of new applications and devices using the network. This is particularly crucial for the design and implementation of future DC networks and edge computing interconnects envisioning a DC infrastructure, in which all computing (CPU, GPU, TPU, photonic computing), memory and storage elements can communicate through the network. This trend, boosted by the need of intelligent and efficient interconnected computing and heterogenous AI and ML-based workload applications in DC and edge nodes, is therefore opening several challenges to be urgently addressed by the scientific community. This includes a huge capacity demand at reduced cost, power consumption and footprint, the need for efficient resource utilization with improved flexibility and scalability, transparency and low latency protocol, just to name some of them.
So, how will the future DC infrastructure be designed to support hyperconnectivity? Aim of the workshop is to explore photonic technologies and network architectures for enabling the next generation DC infrastructure and edge computing, with focus on machine learning (photonic neural networks) applications, co-package optics and flexible Tb/s capacity solutions. Furthermore, innovative architectures based on fiber to the servers and disaggregation/composition of IT resources will be also discussed with special focus on their scalability and requirements.
Some of the key questions that will be discussed by industry and academic panelists, considering ultimately cost and power consumption, are:
- How to scale the DC infrastructures while enabling power-efficient intelligent computing?
- What is the required server/blade connectivity in next generation DC infrastructures?
- What is next after 100 Tb/s electronic switch to support high capacity and high connectivity networks?
- Which are the promising DCI solutions for flexibly supporting Tb/s capacities?
- Is coherent technology the option to be adopted?
- Can photonic switches (in combination with novel network architectures) play a role?
- What is the optimal interface bandwidth/format of next generation edge nodes?
- Are current technologies mature or an evolution is needed? In what direction?
- Is fiber to the servers a viable solution or it only further exacerbates the network and switches requirements?
- Can disaggregation be practical to efficiently use the IT resources and replace inefficient server centric architectures?
- Is the transparency of the photonic switches a real enabler of low latency protocol for disaggregated architectures?
Speakers:
Ryohei Urata, Google, USA
Datacenter Networks and Interconnect: Challenges and Requirements for the Next Decade
Jiajia Chen, Facebook, USA
Quality of Service Aware Optical Backbone Migration for Mega Data Centers
Chongjin Xie, Alibaba, USA
Optics will Play a Bigger Role in Future DC Infrastructure
Kai Shi, Microsoft, United Kingdom
Towards Power-Efficient Data Center Networks
Andy Bechtolsheim, Arista Networks, USA
Power-Efficient High-speed Datacenter Networks
Rob Sherwood, Intel, USA
The Network Inside Vs. Outside Your Servers
Aaron Zilkie, Rockley Photonics, United Kingdom
High Density and Power Efficient Silicon Photonics for Advanced Datacenter Architectures
Shu Namiki, AIST, Japan
Disaggregated Optical Layer Switching That Turn Servers Inside Out
Paraskevas Bakopoulos, Nvidia, Greece
Opportunities for Photonics in AI Datacenters
Bert-Jan Offrein, IBM Switzerland
Neuromorphic Computing Technologies and the Anticipated Impact on Datacenters
Jose Capmany, iPronics, Spain
Programmable Photonics for Data Center Applications: Edge and Cloud Scenarios