Accelerator Performance Engineer – Gemini

Google Inc.; Mountain View, San Francisco (CA, CA); On-site; USD 147,000 - 210,000; Full time

Job summary

Google DeepMind is seeking a Software Engineer to advance performance for Gemini models on hardware accelerators such as TPUs and GPUs. You will work on 2D/3D test environments, create visualizations, and develop tooling to measure and improve efficiency, spanning from model configurations to compiler passes.

You will profile distributed workloads, implement high-performance kernels, and contribute to end-to-end optimization.

Qualifications

• Bachelor’s degree or equivalent practical experience.
• 2 years of experience with software development in C++ or Python.
• Experience with profiling tools (e.g., gProf, Valgrind, or VTune) for performance optimization.
• Experience in software development for hardware accelerators (e.g., CPUs, GPUs, or TPUs), including memory-hierarchy or instruction-level tuning.

Responsibilities

• Improve training and serving efficiency of Gemini models on hardware accelerators (TPUs/GPUs) across configurations and compiler passes.
• Profile large-scale distributed workloads to diagnose bottlenecks and identify optimization opportunities.
• Design and implement high-performance, low-level custom kernels for model operators to maximize hardware utilization.
• Build and maintain benchmarking pipelines and diagnostic tooling to track, reproduce, and guard against regressions.
• Influence accelerator architectures and compiler roadmaps with empirical workload profiles and model requirements.

Skills

C++

Python

Profiling tools

Hardware accelerators

Education

Bachelor's degree

Tools

gProf

Valgrind

VTune

Job description

Google DeepMind is seeking a Software Engineer to advance performance for Gemini models on hardware accelerators such as TPUs and GPUs. You will work on 2D/3D test environments, create visualizations, and develop tooling to measure and improve efficiency, spanning from model configurations to compiler passes.

You will profile distributed workloads, implement high-performance kernels, and contribute to end-to-end optimization.