Meta MTIA 400

The Meta MTIA 400 is the first MTIA chip Meta designed for raw performance rather than cost efficiency alone, and Meta states it delivers performance competitive with leading commercial products. It evolves the MTIA 300 by combining two compute chiplets to double compute density, and adds support for enhanced MX8 and MX4 low-precision formats that matter for generative AI inference. Against the MTIA 300 it offers 400 percent higher FP8 throughput and 51 percent more HBM bandwidth, reaching 12 PFLOPS of MX4, 6 PFLOPS of FP8 and 3 PFLOPS of BF16, with 288 GB of HBM at 9.2 TB/s in a 1200 W module. Seventy-two MTIA 400 devices connect over a switched backplane to form a single scale-up domain, filling a rack-scale system that Meta cools with air-assisted liquid cooling. That choice is deliberate: while the part also supports facility liquid cooling, AALC lets Meta deploy it rapidly into legacy data centers that were never plumbed for direct liquid cooling. MTIA 400 keeps full support for the ranking and recommendation workloads its predecessor targeted while broadening coverage to general generative AI.