Meta is testing its first in-house AI chips: report
Meta, the parent company of Facebook, Instagram, and WhatsApp, is testing its first in-house chip designed to train artificial intelligence (AI) systems, according to sources cited by Reuters. This move signifies a significant step in Meta's endeavors to decrease its reliance on external suppliers such as Nvidia and diminish its substantial infrastructure costs.
The new chip, forming part of Meta's Meta Training and Inference Accelerator (MTIA) series, is a dedicated accelerator customized for AI-specific tasks. This specialization potentially makes it more power-efficient than the general-purpose graphics processing units (GPUs) presently utilized for AI workloads.
Collaboration and Production
Meta is collaborating with Taiwan Semiconductor Manufacturing Company (TSMC) for chip production, with a small-scale deployment already in progress. A successful outcome would lead the company to scale up production for broader utilization.
Cost Management and Vision
Developing in-house chips is intrinsic to Meta's strategy to manage its escalating expenses, which are forecasted to range from $114 billion to $119 billion in 2025. This projection includes up to $65 billion in capital expenditure driven by AI infrastructure.
Future Plans
The company aims to employ its proprietary chips for AI training by 2026, commencing with recommendation systems for Facebook and Instagram feeds, before extending to generative AI products like its Meta AI chatbot.
Challenges and Learnings
Meta's journey in chip development has encountered hurdles. Previously, the company abandoned an in-house inference chip following an unsuccessful test deployment in 2022, leading to a return to Nvidia GPUs. Despite this setback, Meta remains one of Nvidia's major customers, heavily relying on its GPUs for AI model training and inference tasks.
Industry Trends and Concerns
The push for custom silicon comes amidst growing doubts in the AI research community regarding the sustainability of expanding large language models through increased data and computing power. These apprehensions were underscored by the recent introduction of cost-effective, computationally efficient models from DeepSeek, which prioritize inference over traditional training methodologies.
Future Prospects
If the chip demonstrates proficiency, it could aid Meta in cost reduction and offer greater control over its AI infrastructure. Presently, the company is navigating a balance between its in-house initiatives and its dependence on Nvidia's predominant GPU technology. Meta has not issued public statements regarding the testing of these in-house AI chips.
