NEAT: Interconnection Networks for Heterogeneous AI Accelerators

Time period:

1 January 2026 – 31 December 2030

Project leader:

Yuan Yao

Funder:

Swedish Research Council

Type of award:

Grant for positions or stipends

Total funding:

4,800,000 SEK

Purpose The rising computational demands of large AI models are driving a need for heterogeneous accelerator systems, integrating different kinds of accelerators—both energy-efficient ones and powerful but energy-intensive ones—together to balance system´s computing capability and energy use. This diversity highlights the importance of efficient data communication among accelerators, which is challenging, as diverse accelerators have diverse characteristics. Existing interconnects—such as Nvidia´s NVLink/NVSwitch for GPUs—fail to address this challenge because they target specific accelerator types and are fixed-function, meaning they lack the flexibility to adapt to heterogeneous systems. This project, NEAT, closes this gap by proposing a novel programmable interconnect that can dynamically adapt to accelerators with various architectures, enabling smooth and energy-efficient data movement among them.Aims and time plan Planned over 4 years, NEAT aims to (i) develop a unified programming interface to bridge architectural differences between accelerators, and (ii) co-optimize system performance and energy for the interconnected accelerators running large AI workloads. NEAT will be evaluated through both system-level simulation and hardware prototyping.Significance NEAT will shift the role of interconnects from passive data conduits to active, programmable infrastructures, laying the foundation for scalable and sustainable AI.