Software Engineer, Performance, YouTube ML Efficiency

San Bruno, CA

Full-time

Onsite

$141K/yr - $202K/yr

Entry, Mid Level

YouTube is a platform that empowers individuals to share their stories and connect with others through technology and creativity. The role involves improving efficiency in machine learning models, collaborating with quality teams, and optimizing TPU hardware for performance enhancements.

Apply Now

Responsibilities

Redesign components to improve efficiency, including implementing ragged tensors, scaling embedding tables, and developing new optimizers to change how models utilize resources
Partner with quality teams to enable foundational model changes. Identify and unblock performance detractors, allowing experimenters to scale model complexity and training procedures that were previously resource-prohibitive
Act as a local expert on TPU hardware. Push hardware limits by offloading logic to SparseCore, guiding model evaluation based on hardware constraints, and writing low-level Pallas kernels for maximum throughput
Address issues ranging from low-level infrastructure improvements and accelerated linear algebra (XLA) compiler tuning to leveraging application-level insights like traffic patterns and request deduping
Design self-tuning frameworks and analytics to automate performance tuning and resource attribution

Qualification

Required

Bachelor's degree or equivalent practical experience
2 years of experience with software development in one or more programming languages
2 years of experience with computer architecture, performance analysis, and performance modeling
2 years of experience with performance, systems data analysis, visualization tools, or debugging

Preferred

Master's degree or PhD in computer science or related technical fields
2 years of experience with data structures and algorithms
Experience developing accessible technologies
Experience with GPU or TPU performance analysis
Experience with ML frameworks such as TensorFlow, JAX, and PyTorch, or ML compilers (e.g., accelerated linear algebra (XLA))