Software Development Engineer, Neuron Collectives, Annapurna Labs

About the role

The AWS Neuron Collectives team is seeking a Software Engineer to optimize collective operations for AWS Trainium. Trainium is one of Amazon's highest priority initiatives, powering the frontier AI models being trained today. Collectives are the critical operations that scale AI compute across the data center. You'll work in depth to optimize compute for the specific topologies used to train modern LLMs.

Responsibilities

• Enhance collective algorithms and topologies for optimal training performance
• Use tools like Neuron Explorer to identify bottlenecks in compute and bus bandwidth utilization
• Monitor and analyze processor, DMA, firmware, and workload metrics
• Optimize collective operations to scale AI compute across the data center
• Work closely with the hardware team to co-optimize software and Trainium silicon
• Develop and optimize C/C++ implementations of collective communication patterns
• Investigate and implement improvements for specific training topologies used by modern LLMs
• Build and maintain analysis frameworks and automation solutions

Qualifications

• Experience building complex software systems that have been successfully delivered to customers
• Experience contributing to the architecture and design (architecture, design patterns, reliability and scaling) of new and current systems
• Bachelor's degree in computer science or equivalent
• Knowledge of engineering practices and patterns for the full software/hardware/networks development life cycle, including coding standards, code reviews, source control management, build processes, testing, certification, and livesite operations
• Experience in development in the last 3 years, or experience in embedded development in C/C++
• Master's degree in computer science or equivalent preferred
• Experience with hardware/software integration and real-time systems
• Familiarity with collective communication algorithms (e.g., all-reduce, all-gather) or distributed training frameworks