Meta has introduced MTIA 300, the first training accelerator in its new family of internally developed chips optimized for recommendation and ranking models. The chip represents a significant departure from Meta's previous MTIA generation, which focused on inference tasks, and signals the company's intent to control the full AI training stack — from silicon to networking to software.
The most notable innovation in MTIA 300 is the integration of networking components directly into the processor package. Two chiplets each contain six 800 Gbps RDMA interfaces, providing a combined 1.2 terabytes per second of input/output bandwidth. This approach eliminates the traditional separation between compute and networking, reducing latency and increasing throughput for the large-scale distributed training workloads that power Meta's recommendation systems.
Meta also developed the HCCL (Heterogeneous Compute Communications Library) communications library in parallel with the chip. This allows the chip's hardware and the software that orchestrates communication between chips to be co-optimized — a strategy that mirrors what NVIDIA achieves with its NVLink and NCCL ecosystem, but tailored to Meta's specific workload characteristics.
The move confirms an increasingly clear trend across the technology industry: major AI companies are designing chips, networking, and software together to reduce costs and performance losses that arise from using off-the-shelf components. Google has pursued this with its TPUs and custom networking, and Amazon with its Trainium chips. Meta's entry into this space suggests that the market for custom AI training silicon is expanding beyond cloud providers to include companies that operate massive recommendation and ranking systems.
For the broader industry, MTIA 300 raises the question of how far the vertical integration trend will go. As models grow larger and training workloads more complex, the ability to optimize across the full stack — from transistors to networking protocols to training frameworks — may become a decisive competitive advantage.
