01 Jan 0001
00:00 - 00:00
The unprecedented growth of AI models’ size and complexity made interconnecting XPUs a de facto standard for improving accelerator performance and unlocking a shared pool of memory and compute resources, putting immense pressure on the backend – also known as the scale-up – network, given that it carries ~85% of the data traffic and is today copper-based. To compete with such mature technology, photonic solutions are required to address simultaneously a multitude of challenges, including power consumption, thermal management, reliability and system cost together with large radix, exceptional bandwidth and near-zero latency. This requires a holistic approach to design, where aspects such as packaging and signal integrity play a key role in enabling photonics to reach its peak performance. This workshop aims to explore how photonics can extend the scaling trajectory that copper has enabled across previous interconnect generations and what are the prerequisites for mass-deployment of photonics in the scale-up domain.
The key questions to address in this workshop are:
- What is the target energy efficiency and FIT (failures in time) for photonics to become competitive with copper?
- Which optical sources and modulation technologies hold the potential to address all scale-up requirements?
- Should photonic solutions for scale-up strive to be resilient to harsh thermal environment of XPUs or is thermally stable environment an imperative for adopting photonics?
- What are the design requirements allowing DSP to be eliminated?
- What are the requirements for packaging density, physical dimensions and cost/bit for optics in the scale-up network?
- What are the gaps in the supply chain that limit large-scale photonics deployment in scale-up?
- Which emerging photonic technologies or disruptive system breakthroughs are most likely to reshape the future trajectory of scale-up networks?
Organizers
-
Angelina Totovic
Celestial AI, Greece
-
Connie Chang-Hasnain
Berxel Photonics, USA
-
Mahdi Nikdast
Colorado State University, USA
-
Matthew Traverso
Broadcom, USA
-
Paraskevas Bakopoulos
NVIDIA, Greece