Designing the network of tomorrow’s ML systems

Project ID: 2531ad1512

(You will need this ID for your application)

Research Theme: Artificial Intelligence and Robotics

Project Summary:

Are you passionate about machine learning and eager to tackle the challenges of scalable, energy-efficient computing? Join our cutting-edge research project aimed at revolutionizing the infrastructure behind advanced machine learning models.

The Challenge

Training state-of-the-art models like Google’s PaLM demands immense computational resources—up to 3.2 million kilowatt-hours (kWh) of electricity, enough to power 400,000 UK households for a day. As models grow larger and more complex, traditional electronic networks are hitting their limits in bandwidth and energy efficiency.

A new networked ML paradigm?

Optical switches present a promising alternative. Leveraging photons instead of electrons, they offer up to 100 times more bandwidth and lower latency while consuming up to 50% less energy per bit. However, their point-to-point nature and lack of co-design with machine learning collectives have limited their effectiveness.

Our research bridges this gap by optimizing network topologies for optical networks in ML workloads. By redesigning ML collectives to align with optical networks’ point-to-point nature—using pipelined ring or tree-based topologies—we reduce communication overhead, enhance scalability, and accelerate training times.

Why Join Us?

Innovate at the Forefront: Engage in groundbreaking research addressing the technological and environmental challenges of machine learning’s growing demands. Industrial Collaboration: Leverage our partnerships with Google, Meta, and Broadcom to gain unique insights and influence future ML and network designs. Make an Impact: Preliminary studies show co-designed optical switches can reduce communication time by up to 40% and energy use by 30%. Contribute to faster, more efficient, sustainable computing infrastructures. Who Should Apply?

We seek motivated candidates with a strong background in computer science, electrical engineering, or related fields, interested in machine learning, networking, and sustainable computing.