Meta MTIA 200

MTIA 200 is the same 8x8 grid of processing elements as MTIA 100 rebuilt on TSMC 5nm and clocked at 1.35 GHz instead of 800 MHz, inside a 90 W envelope instead of 25 W. The result is 354 TOPS of dense INT8 and 177 TFLOPS of dense FP16 or BF16, roughly three and a half times the first generation, with the memory grown to match: 256 MB of on-chip SRAM at 2.7 TB/s and 128 GB of LPDDR5 at 204.8 GB/s. It is also the generation where Meta started labelling sparsity, publishing 708 TOPS of INT8 and 354 TFLOPS of FP16 with structured sparsity alongside the dense pair, which is why the dense and sparse columns on this row are both real published numbers rather than one of them being derived. The deployment shape changed too: rather than twelve dual M.2 cards in a general-purpose server, MTIA 200 ships in a rack of up to 72 accelerators, three chassis of twelve boards carrying two chips each. Like every MTIA generation it is captive to Meta and cannot be bought or rented.