Particle.news
Download on the App Store

Google Pursues 'TorchTPU' to Align TPUs With PyTorch in Reported Meta Collaboration

The push targets the CUDA lock-in that keeps PyTorch workloads on Nvidia GPUs.

Overview

  • Google is developing an initiative called TorchTPU that aims to make TPUs work more seamlessly with PyTorch, according to Reuters sources.
  • Google is considering open-sourcing parts of the software to speed adoption, though details and timing remain unconfirmed.
  • A Google Cloud spokesperson declined to discuss specifics but said the goal is customer choice and noted accelerating demand for both TPU and GPU infrastructure.
  • Google is working closely with Meta on the effort and has discussed expanded TPU access for Meta, Reuters reported, after earlier disclosures by The Information.
  • A long-standing Jax/XLA focus has hindered external TPU uptake, and Google has recently widened availability by selling TPUs into customer data centers and appointing Amin Vahdat to lead AI infrastructure.