Building the software layer for neuromorphic silicon

Founding story

The gap was the software, not the silicon

In 2021, Andrei Volkov was finishing a PhD in computational neuroscience and had spent the previous several years in chip design at an embedded silicon company in the Bay Area, working on low-power processors for sensor applications. He had watched the neuromorphic silicon ecosystem mature — energy efficiency numbers that no conventional architecture could match — and kept running into the same problem: there was no standard way to take a trained model and compile it to run on any of these chips.

Every team working with neuromorphic hardware was building their own compiler passes, their own runtime, their own calibration tooling — from scratch, in parallel, incompatibly. The constraint wasn't access to good silicon. It was the missing infrastructure layer that sat between PyTorch and the chip.

Volkov co-founded Neurmorph to build that layer: a compiler and runtime that targets any supported neuromorphic silicon from standard PyTorch, without requiring the engineering team to understand SNN training theory or ISA-level chip details. Neurmorph is a software stack. We do not design or manufacture neuromorphic silicon — we build the tooling that makes existing and future silicon deployable at scale.