ML Compiler Engineer, Neuron Kernel Interface

AWS Machine Learning accelerators are at the forefront of AWS innovation. Trainium delivers best-in-class ML training performance with the most teraflops (TFLOPS) of compute power for ML in the cloud. This is all enabled by the AWS Neuron Software Development Kit (SDK), which includes an ML compiler, the Neuron Kernel Interface (NKI) compiler, and a runtime that natively integrates into popular ML frameworks such as PyTorch and JAX. Neuron Kernel Interface (NKI) is a bare-metal language and compiler for directly programming AWS Trainium instances. You can use NKI to develop, optimize, and run new operators directly on hardware while making full use of available compute and memory resources. Explore NKI: - https://awsdocs-neuron.readthedocs-hosted.com/en/latest/nki/index.html AWS Neuron is used at scale by customers such as Epic Games, Snap, Airbnb, Autodesk, Amazon Alexa, and Amazon Rekognition, along with many others across a range of segments. Amazon's Annapurna Labs is responsible for building innovative silicon and software for AWS customers. We are at the forefront of innovation, combining cloud scale with the world's most talented engineers. Our team covers multiple disciplines including silicon engineering, hardware design and verification, software, and operations. With such breadth of talent, there is opportunity to learn all the time. We operate in spaces that are very large, yet our teams remain small and agile. There is no blueprint. We're inventing. We're experimenting. When you couple that with the ability to work on so many different products and services, it makes for a unique learning culture. Learn more about our history: - https://www.amazon.science/how-silicon-innovation-became-the-secret-sauce-behind-awss-success Explore the Product: - https://aws.amazon.com/machine-learning/neuron/ As an ML Compiler Engineer with the Annapurna Neuron team, you develop and scale compilers to handle the world's largest ML workloads on AWS Trainium. You design and ship features across the compiler stack, work directly with Amazon ML partner teams to unblock their workloads, and grow your skills alongside senior engineers. You want to work close to the hardware, help bring new products and features to market, and ship high-quality code. A background in Machine Learning, AI accelerators, or Kernel performance is helpful, but not required. In order to be considered for this role, candidates must be located near or willing to relocate to Seattle or Cupertino. Key job responsibilities - Design and build compiler features for the Neuron compilers. Your work may span the frontend, intermediate representations, optimization passes, and code generation for the hardware. - Own features from design through production: design, implement, test on hardware and simulation, and support customers. - Work with Amazon ML teams (PyTorch, JAX, and internal model teams) to understand their workloads and unblock them. - Give hardware teams compiler and programmability feedback on upcoming chips. - Review code and designs, and mentor newer engineers. A day in the life You start the day reviewing a teammate's code for a new compiler feature, leaving feedback before it merges. Mid-morning, you work with a model team whose kernel isn't reaching peak performance on Trainium, and you trace the bottleneck through the compiler's passes. After lunch, you write and test the feature you designed last week, then run it on hardware to confirm correctness. The team moves quickly and you are always in the details. About the team We are the compiler teams behind AWS Neuron, the software that runs machine learning models on Amazon Trainium. Our ML compiler takes a full model from PyTorch or JAX and compiles the graph to run across the chip. The Neuron Kernel Interface (NKI) is a bare-metal language and compiler that lets customers program Trainium directly, down to the compute and memory, for a new operator or a faster kernel. Together we own the path from framework model to silicon. Our mission is to make Trainium fast and fully programmable. Our teams are small and agile, and there is opportunity to learn all the time. We work closely with the hardware, runtime and framework teams, and with the Amazon ML teams who use our compilers.