Meta
Building the next evolution of digital connection.
Start new thread
Zac Zuo

MTIA 300 - Meta's 3rd-gen custom AI chips for GenAI inference

by
Flowtica Scribe
•
Meta is accelerating its custom silicon roadmap with four new MTIA chips in two years. Built with an inference-first focus and native PyTorch integration, they are designed to cost-effectively power GenAI at a massive consumer scale.

Add a comment

Replies

Best
Zac Zuo
Hunter
📌
Hi everyone! Even though Meta is still one of NVIDIA’s biggest customers, they had already been going all-in on their own silicon — and will clearly continue to do so. Meta is explicitly going inference-first instead of building only for giant pretraining jobs, and they can now ship a new MTIA chip (300 → 500) roughly every six months using modular chiplets and a reusable rack design. That is a very different posture from the usual multi-year silicon cycle.