The bottleneck in training recommendation models is not floating-point throughput, but the speed at which data moves between accelerators. Meta addresses this structural issue with MTIA 300, the first chip in its in-house family designed specifically for training ranking and recommendation models, where embedding tables can constitute over 99% of total parameters.

Add AlexTech.ai asPreferred Source on Google

Network Integrated in the Chip Package

Unlike traditional GPU architectures where network interfaces compete for resources with compute kernels or require CPU mediation, MTIA 300 integrates two network chiplets directly into the package. These contain twelve custom 800 Gbps RDMA NICs, providing a total of 1.2 TB/s of I/O bandwidth that never crosses the PCIe bus. This configuration eliminates host-device-NIC overhead and allows dynamic partitioning of bandwidth between scale-up (up to 1 TB/s within a rack) and scale-out (200 GB/s across racks).

All You Need to Know About Meta’s New AI Chip MTIA — https://encord.com/blog/meta-ai-chip-mtia-explained/

Communication as a First-Class Citizen

The core innovation lies in 16 dedicated message engines (MEs), equipped with RISC-V cores and near-memory compute blocks. These autonomously handle collective operations (AllReduce, AllToAll) without touching the 12x6 PE compute grid. The result is near-perfect isolation: simultaneous execution of GEMMs and collectives introduces less than 0.5% degradation in compute throughput, compared to over 20% typical of GPUs. The HCCL library compiles operations into subgraphs executed autonomously by the device, freeing the CPU after initial dispatch.

Production Performance and Roadmap

On a 150-billion parameter recommendation model distributed across 40 accelerators, MTIA 300 achieves communication speeds 3.9 times faster than an equivalent GPU cluster. The 216 GB of HBM3E memory allows larger local batches, reducing communication needs. Meta indicates that the architectural principles of MTIA 300 and HCCL, presented in ISCA '26 and SC26 papers, form the foundation for future generations of its AI silicon.

Meta MTIA 300 | Awesome Agents — https://awesomeagents.ai/hardware/meta-mtia-300/

Strategic Context

This launch fits into a broader hardware independence strategy. As reported by AlexTech, Meta invested $6.5 billion with Samsung Foundry for 2nm production. The MTIA 300 chip joins the Iris project and aligns with the wider trend of custom AI chips reshaping the AI infrastructure market.