Particle.news

Google Explores Marvell Partnership for AI Memory Chip and Inference TPU

The unconfirmed talks point to Google's drive to diversify TPU suppliers to curb reliance on Nvidia GPUs.

Overview

  • The Information reported Sunday that Google is discussing two chips with Marvell, a memory processing unit to work alongside Tensor Processing Units and a new TPU tuned for running models after training.
  • Marvell shares rose about 6% to 7% in premarket trading Monday after the report, then pared gains as investors reassessed the uncertainty around any deal.
  • JPMorgan said claims that Marvell had won Google’s TPU business were false, adding that Broadcom remains the lead partner on TPU design work.
  • Google and Marvell have not commented, with the report saying the memory chip design could be finished next year before moving to test production.
  • The plan targets two pain points in deployed AI by offloading data movement to a dedicated memory chip and speeding inference, which could cut costs for users of services like chatbots and search if Google scales TPUs more broadly.