Trainium (NKI) Kernel Expert

What You'll Do7

• Review NKI kernel development tasks for correctness, hardware fit, and performance quality.
• Assess CUDA to NKI migration efforts for fidelity to the original kernel's behavior and performance.
• Identify Trainium-specific bottlenecks such as NeuronCore pipeline utilization, tensor-engine throughput, and memory-bandwidth constraints.
• Provide rubric-based written feedback that documents strengths, issues, and required changes.
• Verify cross-platform numerical correctness, including accumulation order, rounding behavior, and mixed-precision semantics between GPU and Trainium.
• Evaluate whether tasks respect NKI patterns like tile-based computation, SBUF/PSUM/HBM memory hierarchy, partition-dimension constraints, and DMA orchestration.
• Collaborate with the AI lab to ensure kernel tasks align with Trainium hardware capabilities.

Requirements9

• 2+ years of hands-on kernel development or tuning with NKI on AWS Trainium or Inferentia2.
• Deep knowledge of NKI patterns: tile-based computation, SBUF/PSUM/HBM memory management, partition dimension constraints, and DMA orchestration.
• Proven experience evaluating CUDA to NKI migration quality.
• Familiarity with Trainium performance profiling, including NeuronCore pipeline utilization, tensor-engine throughput, and memory-bandwidth bottlenecks.
• Experience defining or evaluating cross-platform numerical correctness standards, including mixed-precision semantics.
• Experience with AWS Neuron SDK, Neuron Compiler internals, or NKI kernel libraries (preferred).
• Prior CUDA or Triton kernel development (preferred).
• Knowledge of Trainium hardware specifications such as NeuronCore-v2 architecture, on-chip SRAM, and supported data types FP32, BF16, FP8, INT8 (preferred).
• Experience benchmarking ML training workloads on Trn1 or Trn2 instances (preferred).

Who Should Apply

The right expert has spent at least two years writing and tuning NKI kernels on AWS Trainium or Inferentia2. They can look at a CUDA kernel and judge whether its NKI translation preserves both accuracy and performance. This role is less suitable for CUDA or Triton developers who have not worked with NKI, because the review work centers on Trainium-specific patterns. Candidates often get rejected when they list NKI exposure but cannot describe SBUF/PSUM management or DMA orchestration in detail. Another common miss is lacking a method for checking cross-platform numerical correctness between GPU and Trainium.

Salary Insight