NKI Kernel Optimization Specialist

Obsidian; San Francisco (CA); On-site; USD 150,000 - 210,000; Full time

Job summary

Obsidian is seeking a specialist to evaluate the quality, correctness, and hardware-appropriateness of Neuron Kernel Interface (NKI) development tasks used to train and evaluate frontier AI models.

You will assess CUDA→NKI migration fidelity, Trainium-specific performance insights, and cross-platform numerical-correctness standards, delivering rubric-based written feedback.

Qualifications

• Hands-on kernel development or optimization with NKI on Trainium/Inferentia2 hardware.
• Understanding of tile-based computation and Trainium memory hierarchy.
• Experience assessing CUDA→NKI migrations.
• Familiarity with cross-platform numerical correctness standards.

Skills

NKI kernel development

Trainium hardware

Tile-based computation

Memory hierarchy (SBUF/PSUM/HBM)

DMA orchestration

CUDA→NKI migration

Trainium profiling

Numerical correctness standards

Job description

Evaluate the quality, correctness, and hardware-appropriateness of Neuron Kernel Interface (NKI) development tasks used to train and evaluate a frontier AI lab's models. You'll assess CUDA→NKI migration fidelity, Trainium-specific performance-optimization quality, and cross-platform numerical-correctness standards — and provide clear, rubric-based written feedback.

Basic Qualifications

• 2+ years of hands-on experience developing or optimizing kernels using the Neuron Kernel Interface (NKI) targeting AWS Trainium/Inferentia2 hardware
• Strong understanding of NKI-specific development patterns: tile-based computation, SBUF/PSUM/HBM memory-hierarchy management, partition-dimension constraints, and DMA orchestration
• Demonstrated experience assessing CUDA→NKI migration quality
• Familiarity with Trainium-specific performance profiling (NeuronCore pipeline utilization, tensor-engine throughput, memory-bandwidth bottlenecks)
• Experience defining or evaluating cross-platform numerical-correctness standards (GPU vs Trainium accumulation order, rounding behavior, mixed-precision semantics)

Preferred Qualifications

• Direct experience with AWS Neuron SDK, Neuron Compiler internals, or contributions to NKI kernel libraries
• Prior CUDA or Triton kernel development
• Familiarity with Trainium hardware specifications (NeuronCore-v2 architecture, on-chip SRAM topology, supported data types: FP32/BF16/FP8/INT8)
• Experience benchmarking ML training workloads on Trn1/Trn2 instances