Meta and Broadcom Revolutionize AI with Four New MTIA Chips

Meta has recently unveiled four new generations of custom AI chips, named MTIA 300, 400, 450, and 500. Developed in partnership with Broadcom, these chips are designed to make artificial intelligence more cost-effective to deploy on Meta's platforms, thereby reaching billions of users.

The MTIA chips follow a rapid development cycle, with approximately six months between each generation. Between the MTIA 300 and the MTIA 500, memory bandwidth (HBM) has increased by 4.5 times, while computing power has been boosted by 25 times.

Chip Generation Details

The first of these chips, the MTIA 300, is optimized for ranking and recommendation (R&R) models and is already in production. The MTIA 400, on the other hand, is the first to compete with commercial products in terms of raw performance. A set of 72 MTIA 400 chips constitutes a scaling domain, and this generation has already completed its lab tests and is being deployed in data centers.

The MTIA 450 and 500 are specifically designed for generative AI inference. The MTIA 450 doubles the HBM bandwidth compared to the MTIA 400, thus surpassing current commercial products. These chips support low-precision data formats, such as MX4 and MX8, reducing the computational power required for inference without compromising model quality. The MTIA 500 goes even further by adding 50% more HBM bandwidth and up to 80% additional HBM capacity. Mass production of these two chips is planned for 2027.

Performance and Software Compatibility

The MTIA chips have been designed around well-established industry standards, such as PyTorch, vLLM, and Triton. This allows developers to port existing models to MTIA without requiring special adaptations and to run them on both GPUs and MTIA. Additional technical information is available on Meta's blog.

Meanwhile, Meta continues to collaborate with AMD and Nvidia for GPUs. In February 2026, Meta announced a billion-dollar contract with AMD to provide up to six gigawatts of AMD Instinct GPU computing power dedicated to Meta's AI workloads.