Meta plans to start producing its own artificial intelligence chip this fall, which will allow it to significantly increase its computing power. The new data center chip, developed by the company itself, is codenamed Iris.
Meta plans to start producing its own artificial intelligence chip this fall, which will allow it to significantly increase its computing power. The new data center chip, developed by the company itself, is codenamed Iris.
According to an internal company document seen by Reuters, Iris is part of a larger Meta Training and Inference Accelerators (MTIA) project that will see the release of four generations of processors. The specialized hardware will be used to improve AI algorithms on social networks Facebook and Instagram.
According to an internal memo, the chip was tested for just six weeks and no major issues were identified during that time. Broadcom is the design partner, while Taiwan Semiconductor Manufacturing Co (TSMC) will be responsible for direct manufacturing.
The main goal of the development is to supplement the massive volumes of GPUs that Meta currently purchases from Nvidia and AMD. Implementing third-party solutions on such a scale has proven to be a difficult task, requiring significant time and money. In-house production will allow Meta to significantly reduce computing resource costs and become more independent of hardware suppliers on the market.
The memorandum also outlines a plan to expand the infrastructure in two phases: 7 gigawatts of computing capacity will be put into operation in 2026, increasing to 14 gigawatts by 2027. Meta's projected spending on artificial intelligence infrastructure this year is $145 billion.
The Iris production schedule fits into a broader strategy for developing specialized chips that Meta formally agreed with Broadcom earlier this year, when the two companies agreed to expand their partnership in the field of developing specialized chips for artificial intelligence through 2029.
This isn't Broadcom's only chip collaboration. OpenAI and Broadcom previously developed their own Jalapeño AI processor to eliminate their dependence on Nvidia.
