Leuven | More than two weeks ago
State-of-the-art AI training algorithms rely on massive GPU clusters, creating significant challenges in scalability, power consumption, and cost. Efforts to address these challenges include improving training algorithms, expanding semiconductor fabrication for GPUs and HBMs, investing in nuclear power, and increasing infrastructure funding. However, the major bottleneck today lies in overcoming communication barriers.
Large-scale AI models, such as LLMs, rely on 3D parallelism, which demands high communication overhead. GPUs within a compute node are currently interconnected using electrical technologies (e.g., NVLink, HBI), but scaling up within a package or wafer inevitably requires optical interconnects. Although silicon photonics has matured as a viable solution, current research mainly focuses on point-to-point optical links rather than addressing data movement patterns and 3D parallelism.
This PhD research will explore on-wafer and in-package photonic network architectures tailored for large GPU clusters. The goal is to develop novel multipoint-to-multipoint photonic interconnects optimized for AI workloads involving a massive number of GPUs (e.g., 1 million) and Memory chiplets.
To achieve an energy and cost-efficient AI infrastructure, we will:
- Explore novel photonic network topologies and evaluate their suitability for GPU clusters.
- Develop and assess energy- and cost-efficient chip-to-chip data transmission in silicon photonics inspired by traditional fiber optic networks.
- Analyze the impact of silicon photonic multipoint networking on GPU cluster performance, scalability, and cost-effectiveness.
This position offers the opportunity to work at the forefront of AI, photonic interconnects, and high-performance computing, collaborating with leading experts in both photonics and AI infrastructure.
If you are interested in tackling one of the most critical bottlenecks in AI hardware and shaping the future of large-scale computing, we encourage you to apply!
Required background: Electrical engineering, Physics, Photonics, or related
Type of work: 10% literature study, 50% modelling and design, 40% characterization
Supervisor: Peter Ossieur
Co-supervisor: Joonyoung Kim
Daily advisor: Joonyoung Kim
The reference code for this position is 2025-018. Mention this reference code on your application form.