AI hardware recruitment
for the teams building the silicon.

The profile is narrow

Teams working on AI hardware recruitment problems tend to describe the same difficulty. The role sits between disciplines. An accelerator needs architects who understand how a neural network maps onto a memory hierarchy, compiler engineers who can feed it, and hardware designers who can close timing on the result. A candidate with a strong machine learning background rarely knows what a systolic array costs in area, and a strong digital designer may never have looked at a quantised model.

Because the overlap is small, the people who have it are usually well looked after by their current employers. They are found by following the work, by learning which teams shipped what, and by understanding who built which part of it.

A typical search

A start-up in the Bristol area was building an edge inference chip for cameras and had a working prototype. The founders needed an NPU architect to decide how the next generation would handle transformer workloads. They had tried posting the role and received applications from data scientists, GPU software engineers and CPU architects, each of whom was good and none of whom had designed a dataflow engine.

Joel started by asking the founders which accelerator designs they respected, and why. That gave us a map of the companies where the relevant experience would sit, including a few IP vendors and one university spin-out. We contacted architects there with a plain description of the product, the funding position and the technical decision that was waiting for the hire.

One candidate, who had spent years on a neural processing unit at an IP vendor, was curious about the transformer question because her current roadmap had moved away from it. She spoke to the CTO for an afternoon and asked for sight of the memory bandwidth analysis before agreeing to anything. Once she had seen it, she accepted. The founders told us afterwards that the interview had worked as a design review.

Compiler and software depth

An AI product needs software before it can ship, so we also recruit the compiler, runtime and driver engineers who make an accelerator usable, and we include them in the same search as the hardware team.