Meta, the parent company of Facebook, recently shared new insights into its artificial intelligence (AI) endeavors through a series of blog posts on May 18.
Advancements in AI Technologies
In 2020, Meta developed its initial-generation silicon chip called MTIA (Meta Training and Inference Accelerator) for AI models. This chip was designed to complement GPUs, and Meta utilized it to enhance recommendations. Presently, the company is working on next-generation versions of the chip with improved specifications.
Key takeaways
- Meta disclosed details of its AI infrastructure work in a series of blog posts on 18 May.
- MTIA, the Meta Training and Inference Accelerator, is Meta's own silicon developed in 2020 to complement GPUs, initially used to improve recommendations; next-generation versions are in development.
- The company introduced an AI-optimised data centre design covering both training and inference, working alongside its Meta Scalable Video Processor video transcoding chip.
- Meta announced plans for a 16,000-GPU supercomputer for AI research, expected to rank among the fastest in the world.
- Its LLaMA language model had reached the open-source community in March through an unintended release.
Meta has also introduced an AI-optimized data center design that will be instrumental in AI applications, encompassing both inference and training. This data center design will work in conjunction with Meta’s video transcoding chip, known as Meta Scalable Video Processor (MSVP).
Furthermore, Meta announced its ambitious project to construct a 16,000 GPU supercomputer dedicated to AI research. This supercomputer is anticipated to rank among the fastest AI supercomputers globally.
Limited Disclosures on AI Efforts
Although Meta has not provided an extensive breakdown of its AI technology, the company has previously unveiled information about AI products geared towards the general public.
Throughout this year, Meta and its executives have outlined plans for AI-powered advertising and discussed the implementation of AI in content discovery and user support.
In an unintentional release, Meta’s AI language model, LLaMA, became accessible to the open-source community in March, contrary to the company’s intentions. This leak granted public access to the technology, broadening its reach beyond Meta’s original scope.
Frequently asked questions
Why would Meta design its own AI chips?
Custom silicon can be tuned to the specific workloads a company actually runs, which improves efficiency and reduces dependence on external GPU supply. MTIA was built to complement GPUs rather than replace them.
What is the difference between training and inference here?
Training builds the model; inference runs it to produce results. Meta's data centre design targets both, because the two have different hardware and power characteristics.
What happened with LLaMA?
The model became available to the open-source community in March against Meta's intentions, giving the public access to technology the company had scoped more narrowly.
