Source
MTIA 300: Meta’s First Training Chip with Built-in NICs and Communication-Offloading Engines
Meta has introduced the MTIA 300, an AI accelerator designed for training recommendation and ranking models. It uses integrated network interfaces and dedicated engines to prioritize communication between accelerators.
Why it matters
This efficiency helps Meta train the models that personalize content for app users, overcoming hardware bottlenecks found in general-purpose GPUs.
The details
- Built-in network interfaces provide 1.2 TB/s of total data transfer bandwidth.
- Dedicated message engines handle communication independently from the computation hardware.
- Communication is 3.9 times faster than equivalent GPU clusters for specific models.
Get the weekly recap
The stories like this one, picked and explained — once a week, straight to your inbox.