Meta MTIA 100

MTIA 100 is the chip Meta designed in 2020 and revealed in 2023, the first silicon it built for its own ranking and recommendation models rather than buying. It is a small, deliberately unglamorous part: 64 RISC-V processing elements in an 8x8 mesh on TSMC 7nm, running at 800 MHz inside a 25 W thermal envelope, rated at 102.4 TOPS of INT8 and 51.2 TFLOPS of FP16. What makes it interesting is the memory balance rather than the compute. Recommendation models are bound by embedding lookups, not by matrix throughput, so Meta gave the chip 128 MB of on-chip SRAM at 800 GB/s next to only 64 GB of LPDDR5 at 176 GB/s, which is the opposite of the ratio a training GPU carries. The accelerators mount on dual M.2 boards, twelve to an Open Compute Yosemite V3 server, connected to the host over PCIe Gen4 x8 and to each other through PCIe switches. Meta never sold it, never rented it and never put it in a cloud; it exists to take inference load off the GPU fleet so those GPUs can do something else.