Meta Targets 4.5x Bandwidth with MTIA 300–500 Chips by 2027
Meta is rolling out its MTIA 300 inference chip and has lab-tested the MTIA 400, with MTIA 450 and 500 slated for mass deployment in 2027 and targeting a 4.5x HBM bandwidth and 25x compute boost. These accelerators are designed to offload inference workloads, complementing rather than replacing Nvidia GPUs.
1. Meta Unveils MTIA Accelerator Roadmap
Meta has expanded its in-house Meta Training and Inference Accelerator (MTIA) lineup, deploying the MTIA 300 for ranking and recommendation workloads and completing lab tests on the MTIA 400 ahead of its data center rollout.
2. Significant Bandwidth and Compute Improvements
The MTIA 300 through MTIA 500 series is engineered to deliver roughly 4.5× higher HBM bandwidth and 25× more compute FLOPS, with new chip launches every six months aligned with Meta’s data center growth.
3. Collaboration and Fabrication Partners
Meta collaborates with Broadcom on chip design elements and uses TSMC for fabrication, ensuring advanced process nodes support the MTIA roadmap’s aggressive performance targets.
4. Implications for Nvidia GPU Demand
While Meta continues to buy tens of billions in Nvidia and AMD off-the-shelf chips, expanding MTIA deployments to handle inference workloads may dampen incremental demand for Nvidia’s inference GPUs.