Particle.news
Download on the App Store
13 ARTICLES
·
4h ago

Google, Marvell In Talks To Co-Develop AI Chips Focused on Inference

The talks signal a push to lower inference costs through a paired TPU–memory design.

AI (Artificial Intelligence) letters are placed on computer motherboard in this illustration taken, June 23, 2023. REUTERS/Dado Ruvic/Illustration/File Photo
Image
Google in talks with Pentagon to deploy AI models for classified use.
Image

Overview

  • Reports Sunday said Google and Marvell are in talks to build two custom AI chips, and neither company has confirmed the discussions.
  • One proposed chip is a memory processing unit that sits next to Google’s Tensor Processing Unit to move data faster, while the other is a TPU tuned for inference, which is the step where a trained model generates answers.
  • The companies aim to lock the memory-chip design as soon as next year before test production, according to The Information’s account of the talks.
  • Coverage frames the effort as part of Google’s multi-supplier plan to scale TPU capacity and reduce reliance on a single vendor, with any Marvell work supplementing ties to Broadcom, MediaTek, TSMC, and Intel.
  • Investors reacted to the report with Marvell shares jumping up to 6.3% and Alphabet edging higher, as analysts saw a move to challenge Nvidia’s grip on AI compute and to drive more Google Cloud revenue from TPUs.