Trainium Kernel Correctness and Performance Reviewer

This role tests whether a kernel solution is both mathematically sound and appropriate for the accelerator that will execute it. You will examine NKI development tasks, trace implementation choices to the Trainium architecture, and explain which results meet the evaluation criteria.

Work you will do

• Check CUDA-to-NKI ports for preserved behavior, unsupported assumptions and changes caused by execution order or numeric precision.
• Review tile layout, partition dimensions, DMA transfers and use of SBUF, PSUM and HBM; identify choices that impair correctness or hardware efficiency.
• Interpret profiling evidence for NeuronCore pipelines, tensor-engine throughput and memory-bandwidth limits, separating measured improvements from unsupported optimization claims.
• Assess the numerical comparison standard across GPU and Trainium implementations, including accumulation order, rounding and mixed-precision behavior.

Required background

• At least 2 years building or optimizing NKI kernels for AWS Trainium or Inferentia2.
• Practical command of NKI tiling, memory hierarchy, partition constraints and DMA, plus experience assessing CUDA-to-NKI migrations.
• Ability to evaluate Trainium profiles and define defensible cross-platform numerical checks.

Preferred background

• AWS Neuron SDK or compiler-internals work, or contributions to NKI kernel libraries.
• CUDA or Triton kernel development; NeuronCore-v2/SRAM architecture and FP32, BF16, FP8 or INT8 familiarity.
• ML-training benchmarks on Trn1 or Trn2.

Deliverables

A documented rubric-based review, concrete correctness findings, and a performance assessment tied to the supplied hardware evidence.

Location and schedule

Remote; United States eligibility applies to this track. Availability of up to 40 hours a week can be discussed. IXO will confirm the actual assignment schedule before work; this is not guaranteed weekly volume.

Pay and working terms

$75 $95/hr USD. IXO will agree the assignment scope, hourly rate, schedule and acceptance criteria before work begins. Work is a paid remote expert engagement. Use only public, licensed or otherwise authorized material. Do not provide confidential employer information, personal data or restricted research.