Automating the Design of Bespoke AI Hardware

The Problem

The current landscape of AI hardware suffers from a underlying problem: poor compute occupancy. Despite peak performance of hardware rising year-on-year, the percentage of the chip performing useful computations is diminishing to below 10%. This is caused by a number of factors: the memory wall, end of Moore's law, etc. Fundamentally, the vast majority of AI chips are sitting underutilised.

The Solution

We believe the future of compute is extremely application-specific processor architectures. Instead of developing a "one-size-fits-all" chip, next generation AI applications will require bespoke processor architectures. To achieve this level of customisation, we need to move beyond the traditional semiconductor design flow, instead bringing automation further up the stack. By addressing the compute occupancy problem at the architectural level, we can unlock the full potential of future generations of technology (photonic, neuromorphic, etc.)