Meta Unveils Next-Gen AI Chip, Doubles Performance to Reduce Reliance on Nvidia GPUs
-
Meta unveiled its second-generation AI inference accelerator, MTIA, which doubles compute and memory bandwidth over the previous version.
-
The new MTIA chip architecture balances compute power, memory bandwidth, and capacity for ranking and recommendation models.
-
Key upgrades include more local storage, on-chip SRAM, and LPDDR5 capacity compared to the first MTIA.
-
Meta co-designed the MTIA software stack to synergize with the new hardware for optimal inference performance.
-
While not drastically reducing reliance on Nvidia GPUs yet, MTIA is another step towards Meta's goal of less dependence on external AI hardware.