Trainium (NKI) Kernel Developer for AI Model Evaluation

Role Overview Help evaluate Neuron Kernel Interface development tasks that support the training and evaluation of advanced AI models. You will assess kernel quality, numerical correctness, CUDA-to-NKI migration fidelity, and whether implementations are well suited to AWS Trainium hardware, then deliver clear written feedback using defined evaluation criteria. Key Responsibilities Evaluate NKI kernel development tasks for quality, correctness, and hardware appropriateness. Review CUDA-to-NKI migrations for fidelity and implementation quality. Assess Trainium-specific performance optimization, including effective use of hardware resources. Evaluate cross-platform numerical-correctness standards, including accumulation order, rounding behavior, and mixed-precision semantics across GPU and Trainium environments. Provide clear, rubric-based written feedback. Qualifications At least 2 years of hands-on experience developing or optimizing NKI kernels for AWS Trainium or Inferentia2 hardware. Strong knowledge of NKI development patterns, including tile-based computation, SBUF, PSUM, and HBM memory-hierarchy management, partition-dimension constraints, and DMA orchestration. 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 numerical-correctness standards across platforms. Preferred Qualifications Experience with the AWS Neuron SDK, Neuron Compiler internals, or NKI kernel-library contributions. Prior CUDA or Triton kernel-development experience. Knowledge of Trainium hardware, including NeuronCore-v2 architecture, on-chip SRAM topology, and FP32, BF16, FP8, and INT8 data types. Experience benchmarking machine-learning training workloads on Trn1 or Trn2 instances. Work Terms Remote role, open to candidates located in the United States. Hourly engagement. Compensation $70 to $90 per hour.