Fathom Computing is engineering high-performance machine learning accelerators, built to run both training and inference workflows for very large scale artificial neural networks (ANNs).

 

Artificial intelligence (AI) is an application layer that runs on hardware. Contemporary devices used for Artificial Intelligence (AI) applications - from the smallest energy efficient system on chip to the largest multi-node supercomputers - are quite similar in general construction. Their functionality is derived from digital logic, implemented with billions of transistors in a common silicon substrate connected by metal wires. These devices have a diverse range of architectures with varying levels of application-specific performance, broadly ranging from general purpose CPUs and GPUs to state-of-the-art ASICs.

 

Generational improvements in machine learning (ML) hardware performance will not arise from further incremental optimizations in ASIC design, but from foundational system architecture changes that dramatically reduce bandwidth bottlenecks between processing cores and memory. Fathom’s hybrid electro-optical architecture focuses innovation precisely in this area - resulting in ML performance improvements not otherwise achievable in traditional electronics-only systems.