Meta Unveils MTIA 300 Training Chip with Built-in NICs
Meta announced MTIA 300, its first in-house training and inference accelerator optimized for ranking and recommendation models. The chip features built-in NIC chiplets and a co-designed communication library, HCCL, to offload communication tasks, delivering superior performance over general-purpose GPUs for training recommendation models.
Sources and evidence
Summary last validated Aug 26, 2026
Reader actions
Report an issue
Use this for an incorrect summary, wrong source, duplicate story, or wrong category. Submissions are private and do not change the story automatically.
Related coverage
- chips
MIT Lincoln Laboratory survey tracks AI accelerator evolution
Supercomputing researchers at MIT Lincoln Laboratory are conducting an ongoing survey that documents the evolution of AI hardware, tracking the latest AI accelerator systems. The effort aims to keep hardware relevant for Lincoln Laboratory staff and sponsors.
- infrastructure
Meta Adds NTS Authenticated Time Service at nts.meta.com
Meta's public time service now supports NTS (Network Time Security, RFC 8915) at nts.meta.com. Packets are authenticated so devices can verify time came from Meta and was not modified in transit. Meta says its NTS servers hold no per-client state, with cookie keys derived rather than stored or replicated, and it has open sourced the implementation.
- chips
NVIDIA DGX Spark 64GB Brings Local AI to Developer Desktops
NVIDIA announced that its DGX Spark will be available this month with 64GB of unified memory from manufacturer partners including Acer. The desktop AI system is aimed at developers running increasingly capable open models and AI agents locally rather than in the cloud.
- security
Meta Engineering Post Addresses Private Processing for Meta AI Glasses
Meta published an engineering blog post titled "Bringing Private Processing to Meta AI Glasses." The post discusses glasses as a form factor for AI assistance that understands personal context, but the available excerpt does not describe what private processing involves or confirm any deployment.
- open source
Meta open-sources Rebalancer assignment-problem solver
Meta has open-sourced Rebalancer, a generic, high-performance library for solving assignment problems. The solver has been used internally for over nine years to handle resource allocation problems across Meta. It separates how an assignment problem is specified, stored in memory, solved, and debugged, a design Meta says is crucial for its flexibility and performance.