Meta mtia scale ai chips for billions

Meta acknowledges in its official blog that AI model evolution outpaces traditional chip development cycles—chips are designed based on predicted workloads, but by the time they reach volume production two years later, workloads may have changed significantly. To address this, Meta adopts an iterative approach: each MTIA generation builds on the previous one, using modular chiplets to shorten deployment cycles.

The MTIA family includes MTIA 300 (already used for ranking and recommendation training), MTIA 400 (targeting GenAI and ranking/recommendation, performance comparable to commercial products), MTIA 450 (optimized for GenAI inference, double HBM bandwidth), and MTIA 500 (50% more HBM bandwidth). These chips are deployed or planned for deployment in 2026–2027.

Meta states that MTIA is a key component of its AI infrastructure strategy but emphasizes maintaining a diverse silicon portfolio, leveraging both internal and external best solutions.

According to Meta's published data, from MTIA 300 to 500, HBM bandwidth increased 4.5x and compute FLOPS increased 25x (from MX8 on MTIA 300 to MX4 on MTIA 500). MTIA 450's HBM bandwidth is "far higher than existing leading commercial products," and it introduces hardware acceleration to alleviate Softmax and FlashAttention bottlenecks.

MTIA 450 supports mixed low-precision computation without software overhead for data type conversion, and introduces custom data types to maintain model quality while boosting FLOPS. MTIA 500 further innovates on low-precision data types beyond the 450.

MTIA 400 is the first MTIA chip to compete with leading commercial products in raw performance, has completed lab testing, and is being deployed to data centers. MTIA 300 is described as a "cost-effective" product.

Meta has deployed hundreds of thousands of MTIA chips and integrated them into multiple internal production models, including testing with large language models like Llama. MTIA 300 was initially optimized for ranking/recommendation inference and is now used for ranking/recommendation training; MTIA 400 extends to generative AI workloads; MTIA 450 and 500 are specifically optimized for GenAI inference.

MTIA 400 uses a 72-accelerator scale domain connected via a switch backplane. MTIA 450 is planned for mass deployment in early 2027, and MTIA 500 in 2027.

Meta notes that MTIA chips are developed in collaboration with Broadcom, and Meta has published two ISCA papers (2023 and 2025) disclosing architectural details of the first two generations.