Case Study: Graphcore scales AI experimentation and simplifies tracking with Weights & Biases

How Graphcore is Supporting the Next Generation of Big Models with the Help of W&B

Graphcore, the inventor of the Intelligence Processing Unit (IPU) designed for next-generation AI workloads, faced significant challenges in tracking and managing the massive number of experiments run on its advanced IPU-POD systems. The company was using manual methods like spreadsheets and internal tools, which were inadequate for scaling experimentation across multiple deployment locations and failed to provide a central place for analysis and team collaboration. This complexity hindered their ability to efficiently diagnose issues and develop state-of-the-art models, creating a critical need for a robust solution.

Weights & Biases provided its AI developer platform to address these scaling challenges. The solution gave Graphcore a centralized source of truth for experiment tracking, enabling deep analysis, easy comparison of results, and improved team collaboration. According to Graphcore's Director of Applications, this integration led to a dramatic 50 to 100 times increase in the number of experiments they could run, allowing them to debug issues faster and scale their systems and software development effectively alongside their growing hardware.

Graphcore

Phil Brown

Director of Applications