Meta unveils more advanced AI chip for faster model training
Meta has announced its latest custom artificial intelligence (AI) chip, designed to boost power and speed up training for its ranking models. The new Meta Training and Inference Accelerator (MTIA), or MTIA v2, is specifically engineered to improve Meta's ranking and recommendation models. This innovation aims to streamline training processes and simplify inference (the actual reasoning tasks). MTIA v2 forms a significant component of Meta's long-range strategy for developing infrastructure to support the integration of AI into its services.
MTIAs are now in production
Meta unveiled MTIA v1 in May 2023 with the primary aim of supplying these chips to data centers. The subsequent iteration of the MTIA chip, MTIA v2, is also projected to cater to data centers, indicating a continued focus on this sector. MTIA v1 was not expected to be released until 2025. However, contrary to initial estimates, Meta has confirmed that both versions of the MTIA chips are currently being manufactured.
New AI chip outperforms its predecessor
The MTIA v2 is a 5nm chip, physically larger and equipped with more processing cores than its predecessor, MTIA v1, which was built on a 7nm process. Despite using more power (90W v/s 25W), the next-gen chip offers more internal memory (256MB v/s 128MB) and operates at a higher average clock speed (1.35GHz up from 800MHz). Preliminary tests from Meta suggest that the next-gen chip outperforms its predecessor, MTIA v1, by three times across four models evaluated by the company.
MTIA v2 to be utilized in training generative AI models
Presently, MTIA primarily facilitates the training of ranking and recommendation algorithms. However, Meta has revealed plans to broaden the MTIA v2's functionalities, aiming to enable it to train generative AI models akin to its LLaMA language models.
Investment in next-generation hardware systems
Meta's ambitions for its custom silicon involve not just compute silicon, but also memory bandwidth, networking, and capacity. The company is also investing in other next-generation hardware systems. This information was shared by Meta in a blog post, highlighting their commitment to advancing technology and improving the efficiency of their operations. The company's objective is to tailor its chip design to seamlessly align with both its existing technological framework and forthcoming advancements in GPUs.
Custom AI chips: A growing trend among tech giants
The trend toward custom AI chips is not exclusive to Meta. Other tech giants like Google, Microsoft, and Amazon have also been creating their own chips to cater to the growing demand for compute power as AI usage increases. This need for powerful chips has escalated to such a level that NVIDIA, currently leading the AI chip market, is valued at $2 trillion.
