What the work involves

You'll spend most of your day reading kernels, not writing them. Tasks arrive as NKI implementations — often ported from a CUDA or Triton original — accompanied by claimed speedups, profiling output, or correctness arguments. Your job is to decide whether the kernel is correct, whether it is actually well-suited to NeuronCore hardware, and whether the accompanying reasoning holds up. That means tracing tile shapes against partition-dimension constraints, checking whether SBUF residency and PSUM accumulation are used the way the architecture expects, spotting DMA patterns that will serialize, and asking whether a reported throughput number came from a configuration that means anything.

A large slice of the work is migration fidelity. A CUDA kernel that leaned on warp-level shuffles, shared-memory bank tricks, or `atomicAdd` does not have a one-to-one NKI translation, and a plausible-looking port frequently changes accumulation order, precision semantics, or edge-case behavior without saying so. You'll be asked to name what changed, whether the change is acceptable, and what tolerance a cross-platform equivalence test should actually use for the dtype involved.

What the screen looks for

The platform interview is AI-led and follow-up heavy. It is calibrated to separate people who have shipped NKI kernels on Trn1/Trn2 from people who have read the documentation. Expect to be asked for specific numbers — SBUF capacity per NeuronCore, the partition-dimension limit, what happens when a tile exceeds it — and to be pushed one level past your first answer. Vague architectural gestures get probed until they break. It also tests evaluation judgment separately from engineering skill: whether you can score a kernel that is correct but poorly suited to the hardware, and whether your written rationale would let another reviewer reproduce your score.

Logistics

  • Remote, US-based; work authorization in the US is a stated requirement.
  • Roughly 40 hours per week — this is a full-time-shaped engagement, not weekend piecework.
  • Largely asynchronous, with review volume and rubric revisions delivered through the platform.
  • $70–$90/hour as observed on the listing; rate placement typically tracks depth of hands-on Neuron experience and is not guaranteed.