Machine Learning Fleet Efficiency: Analyzing and Optimizing Large-Scale Google TPU Systems with ML Productivity Goodput

Arissa Wongpanich † Tayo Oguntebi † Jose Baiocchi Paredes † Yu Emma Wang † Phitchaya Mangpo Phothilimthana † Ritwika Mitra † Zongwei Zhou † Naveen Kumar † Vijay Janapa Reddi †‡ † Google ‡ Harvard University · 2025

Read Machine Learning Fleet Efficiency: Analyzing and Optimizing Large-Scale Google TPU Systems with ML Productivity Goodput by Arissa Wongpanich † Tayo Oguntebi † Jose Baiocchi Paredes † Yu Emma Wang † Phitchaya Mangpo Phothilimthana † Ritwika Mitra † Zongwei Zhou † Naveen Kumar † Vijay Janapa Reddi †‡ † Google ‡ Harvard University (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2502-06982.txt.

Description

Arissa Wongpanich and co-authors measure how well Google's vast TPU fleet turns machine learning work into finished training, and use the lens to find waste, in work from twenty twenty-five. Giant chip fleets now train the biggest models. Raw usage rates hide where time really goes.

Their metric, called ML Productivity Goodput, counts genuinely productive work across the whole stack. The study spans thousands of chips running mixed jobs. Layers from model code to scheduling, not just hardware, bent the numbers hard. Workload variety also defied per job scoring. Applying the metric to a live internal fleet surfaced bottlenecks and guided fixes with measured gains.

Catalog record

CREATORArissa Wongpanich † Tayo Oguntebi † Jose Baiocchi Paredes † Yu Emma Wang † Phitchaya Mangpo Phothilimthana † Ritwika Mitra † Zongwei Zhou † Naveen Kumar † Vijay Janapa Reddi †‡ † Google ‡ Harvard University
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE56 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2502-06982

Browse the full library · About the Tower of Babel