Google Cloud Makes Pathways Runtime for Large-Scale TPU Workloads Generally Available

At a glance

Google Cloud announced that Pathways Runtime, which helps run large-scale jobs on its AI chips called TPUs, is now generally available on GKE. It lets multiple TPU groups be managed from a single controller, so jobs can run more reliably.

Key points

1. 1

   A single program can treat multiple TPU groups spread across thousands of chips as if they were local devices.

2. 2

   If hardware fails, the job resumes from a host memory snapshot in a reduced configuration.

3. 3

   Google Cloud said that over the past 90 days, more than 130 million TPU hours were used this way.

4. 4

   Integration with Slurm clusters is still in preview, so its status differs from the GKE general availability release.

Why it matters

As AI models grow, more computing devices are linked together, and managing them is complex and often causes stoppages. This approach can reduce these problems and save cost and time.