Dr. Elena Marchetti spent the better part of fifteen years designing neural accelerators — the kind that end up as one block inside someone else's system-on-chip, budgeted down until they barely matter. The pattern was always the same: the accelerator worked, and then the power budget, the die area, and the schedule all carved it back.
At a compiler workshop in 2020, she met Dr. Tomas Reyes, who'd been building the software side of the same story — squeezing models onto hardware that was never quite built for them. The two of them kept returning to the same question: what would the silicon look like if it were designed for edge inference from the first floorplan, not retrofitted?
"Edge AI kept getting the leftover silicon. We decided to build the silicon first, and let everything else follow."
In 2021 they founded NeuralPeak. The first year was a single test chip and a lot of late nights; the second year brought NP-1 back from the fab, and it worked the way they'd hoped. Today the company designs, tapes out, and ships its own neural processors — and stays deliberately small enough that the people answering your questions are the ones who drew the floorplan.