Building Digital Brains: A New Programming Method For Automatically Generating Neuromorphic Systems

”Speed and scale” [Nature19] is considered as the two (out of four) most important obstacles to solve to advance brain simulations.

The culprit behind these obstacles is a mismatch between traditional (von Neumann) architectures and how the brain operates, leading to slow, inefficient, and power-hungry solutions.

In this computer architecture project, we aspire to research a new programming method that allows domain scientists to automatically generate – or build – “brains”.

These ”brains” – also known as neuromorphic systems – will be based on emerging reconfigurable technology, which allows for specialization, customization, and silicon plasticity.

Our proposed method builds on High-Level Synthesis and separates the description of neuron/synapse dynamics from the hardware synthesis, which leads to a reduced (minimal) learning curve for domain experts while facilitating high-performance, portable, and power-efficient solutions.

Using emerging techniques and memories such as 3D stacking, eDRAM, and ReRAM, we expect to demonstrate how future brains can be automatically generated and yield between 100x-1000x higher capacity or performance compared to state-of-the-art today.

Our project is driven by use-cases from three (out of many) different research fields that can benefit from our methodology, including neuroscience, machine learning, and physics.