Meta 的客製 AI 晶片將於 9 月上線——它能否最終擺脫對 Nvidia 的依賴？
meta 於 9 月開始生產其 mtia 晶片，以降低 gpu 成本並管理 1450 億美元的資本支出預算。這一將硬體內部化的舉措標誌著其致力於減少對 nvidia 的完全依賴，為其 ai 模型提供動力。
meta is ramping up production of its custom mtia chips this september, utilizing a modular design of chiplets developed with broadcom and manufactured by tsmc. these chips are slated for training ranking models and running inference, specifically targeting workloads that do not require the heavy lifting of training large foundation models. this initiative is a direct response to soaring capital expenditures, forecasted between $125 billion and $145 billion, as meta races to deploy 7 gigawatts of compute for its ai efforts. by bringing chip design in-house, meta follows amazon and google in attempting to reduce dependency on nvidia, though it will continue to spend billions on external gpu suppliers.