Meta has started to deploy the new generation of its MTIA (Meta Training and Inference Accelerator) chips in its data centers, and this can be seen as a significant move to make it less dependent on external chip vendors. This relocation follows the company’s having entered into massive GPU acquisitions with Nvidia and AMD, with a two-fold approach to responding to the skyrocketing AI computing requirements.
MTIA chips are tailored to AI inference first of all – that is, the practice of executing trained models across platforms, including Facebook, Instagram, and WhatsApp. Developing chips that are specific to its recommendation systems and AI capabilities, Meta will reduce its long-term computing expenses and enhance its efficiency.
According to industry analysts, custom silicon may cut inference costs by half when used everywhere. Nonetheless, Meta will not stop training the AI models in commercial GPUs. The approach is a reflection of the efforts of Google and Amazon, where tech giants would like to have more control over their AI systems.
