Meta Platforms Inc. has developed four custom AI chips to drive its internal artificial intelligence workloads. The most advanced of these, the MTIA 500, offers 10 petaflops of performance and supports efficient MX4 data processing. This chip, alongside the MTIA 450, is slated for production in 2027 and is optimized for generative AI inference, featuring specialized circuits to accelerate tasks like FlashAttention for large language models.
The MTIA 300, already in production, focuses on ranking and recommendation models with 1.2 petaflops performance. The new chip lineup leverages modular designs, allowing for seamless integration and faster deployment. Meta utilizes custom compilers and a dedicated communication library to optimize AI models for these processors, significantly boosting performance by minimizing data travel times.