Software Engineer III, Performance

Google•Mountain View, CA

1d•$147,000 - $210,000

About The Position

Google is hiring a Software Engineer III to improve performance for open-source machine learning models running on TPUs. The team aims to demonstrate TPU capabilities in industry competitions and optimize models so Google Cloud customers can apply the resulting recipes to their own models. The role involves writing software, manually optimizing models, turning optimization principles into scalable and automatable solutions—including agentic AI tooling—and improving out-of-the-box performance for third-party customers. Engineers will also analyze performance trends and maintain benchmarks to identify and prevent issues in new models.

Requirements

Bachelor’s degree or equivalent practical experience.
At least 2 years of software development experience in one or more programming languages.
At least 2 years of experience with computer architecture, performance analysis, and performance modeling.
Experience experimenting with AI agents or LLM-powered agentic workflows.
Experience in systems engineering.

Nice To Haves

Master’s degree or PhD in Computer Science or a related technical field.
Experience with low-level accelerator programming, such as CUDA, Triton, or Pallas, and performance tuning techniques.
Experience with machine learning, compiler optimization, code generation, and runtime systems for GPU architectures, such as OpenXLA, MLIR, or Triton.
Debugging experience improving performance of single-mode or distributed multi-mode systems.
Expertise tailoring algorithms and machine learning models to exploit accelerator architecture strengths and minimize weaknesses.
Deep understanding of modern machine learning accelerator architectures, including Nvidia or AMD GPUs, TPUs, Trainium, or Cerebras; memory hierarchies; and performance bottlenecks.

Responsibilities

Write product or system development code.
Manually optimize models using techniques such as sharding, quantization, and custom kernels.
Develop principles from manual optimizations that can be automated through agentic AI, such as compute, communication, scheduling, or quantization cost models.
Turn identified principles and recipes into scalable solutions, for example by collaborating with the compiler team on cost modeling or building agentic tooling for headroom analysis.
Improve out-of-the-box software performance and open-source tooling to create scalable solutions for third-party customers.
Develop trend analysis to identify performance issues in previously unseen models and address them proactively.
Maintain model benchmarks to track performance improvements.