The Annapurna Labs team at Amazon Web Services (AWS) builds AWS Neuron, the software development kit used to accelerate deep learning and GenAI workloads on AWS Trainium, Amazon's custom machine learning accelerator. Neuron includes an ML compiler, runtime, collectives library, and application framework that integrate with PyTorch and JAX, so customers can train frontier-scale models on Trainium without rewriting their stack.
The Distributed Training team is at the forefront of training a wide range of models on AWS's custom ML accelerators, supporting novel architectures while maximizing their training performance. Working across the stack from PyTorch and JAX down to the hardware and software boundary, our engineers build the infrastructure that large-scale training depends on, develop new parallelism and numerics techniques, and tune high-performance kernels for the operations that dominate a training step, so every compute unit is doing useful work on our customers' most demanding workloads. We combine deep hardware knowledge with ML expertise to push the limits of training efficiency at scale.
As part of the broader Neuron organization, our team works across multiple technology layers, from frameworks and kernels through to the compiler, runtime, and collectives teams. This is hardware and software co-design in practice. A single throughput gap rarely sits in one layer, so tracing it means following the problem across the stack, deciding where the fix belongs, and working with the team that owns that layer to land it. We not only optimize current performance but also contribute to future architecture designs, since the gaps we characterize today become requirements for the next generation of Trainium. We work closely with customers to enable their models and ensure they train efficiently. This role offers a rare opportunity to work at the intersection of machine learning, high-performance computing, and distributed systems, where you will help shape the direction of AI acceleration technology.
You will architect and implement business critical features, and mentor a team of experienced engineers. We operate in spaces that are very large, yet our teams remain small and agile. There is no blueprint. We are inventing. We are experimenting. It is a genuinely unique learning culture. The team works directly with customers on model enablement, providing hands-on support and optimization expertise so their training workloads reach the performance they need on AWS ML accelerators. We also collaborate with the open source ecosystem, contributing upstream so integration is seamless and performance holds at scale for customers and developers.
Key job responsibilities
You will lead efforts to optimize distributed training performance on Trainium, with a primary focus on training throughput, model FLOPs utilization, and time to convergence across the Neuron software stack. You will work across PyTorch, JAX, and the Neuron compiler and runtime to enable and tune large-scale training workloads on the latest Trainium instances. You will bring up model architectures that have never run on Trainium, identifying the missing operators, sharding strategies, and numerics needed to train them correctly, and then close the gap between correct and fast. You will own the parallelism strategies these models depend on, spanning data, tensor, pipeline, expert, and context parallelism, and apply reduced-precision formats where they measurably pay off. You will profile end to end to determine whether a workload is bound by compute, memory, collectives, or host overhead, then drive the fix to the layer that owns it, working with compiler, runtime, and collectives engineers to land it. You will translate the performance gaps you characterize into requirements that influence future Trainium architecture, and contribute upstream to the open source frameworks our customers train on.
A day in the life
You will collaborate with a cross-functional team of applied scientists, system engineers, and product managers to deliver state-of-the-art inference capabilities for Generative AI applications. Your work will involve debugging performance issues, optimizing memory usage, and shaping the future of Neuron's stack across Amazon and the Open Source Community. As you design and code solutions to help our team drive efficiencies in software architecture, you’ll create metrics, implement automation and other improvements, and resolve the root cause of software defects.
About the team
Annapurna Labs was a startup company acquired by AWS in 2015, and is now fully integrated. If AWS is an infrastructure company, then think Annapurna Labs as the infrastructure provider of AWS. Our org covers multiple disciplines including silicon engineering, hardware design and verification, software, and operations. AWS Nitro, ENA, EFA, Graviton and F1 EC2 Instances, AWS Neuron, Inferentia and Trainium ML Accelerators, and in storage with scalable NVMe, are some of the products we have delivered, over the last few years.
