Meta MTIA 450

The Meta MTIA 450 is a generative AI inference specialist, and the point where Meta stopped optimising for balance and went after decode throughput. Meta doubled HBM bandwidth from the MTIA 400 to 18.4 TB/s, which it describes as much higher than existing leading commercial products, because memory bandwidth is the factor that most limits generative AI inference performance. HBM capacity stays at 288 GB and module power rises to 1400 W. Compute reaches 21 PFLOPS of MX4, 7 PFLOPS of FP8 and 3.5 PFLOPS of BF16, an increase of 75 percent in MX4 throughput over the MTIA 400. Beyond raw numbers Meta added hardware acceleration aimed at specific inference bottlenecks, easing Softmax and FlashAttention pressure and speeding up mixture-of-experts feed-forward network computation. It also supports mixed low-precision computation without the software overhead that data type conversion normally imposes, and introduces custom Meta data types intended to raise throughput while preserving model quality at minimal cost in chip area. Meta frames the balance plainly: the part delivers six times the MX4 throughput of its own FP16 and BF16, which is a statement about how far inference has moved toward low precision.