The next generation of Meta’s custom AI chips will support generative AI models.
In a blog post Wednesday, Meta said the newer version of its Meta Training and Inference Accelerator (MTIA) chip is “being built with AI in mind.”
Meta’s New AI Chip Will Support GenAI Models
The first iteration of MTIA was unveiled in May 2023, particularly to power the company’s deep-learning recommendation models. Meta said the new chips would more than double the compute and memory bandwidth of the predecessor and would support the company’s new generative AI services and other AI workloads.
The next-gen MTIA will be physically larger in design compared to MTIA v1. The chips will feature 256MB in on-chain memory capacity and 2.7 TB/s in on-chain memory bandwidth compared to the 128MB and 800 GB/s, respectively, on the older generation.
Source: Meta Platforms.
The company said the new chips performed 3x better than v1 across four key models during the early test stage.
Meta views its custom-made chips as a long-term venture. “It’s an investment we expect will grow in the years ahead as the compute requirements to support AI models increase alongside the models’ sophistication,” the company said.
Big Techs Turn to Building Custom Chips
Meta building its AI chips weaves on the growing trend of big tech companies turning to produce their own chips to meet their AI workload needs amid the shortage of chips.
“It [MTIA] is proving to be highly complementary to commercially available GPUs in delivering the optimal mix of performance and efficiency on Meta-specific workloads,” Meta said.
Meta’s rivals, including Google and Microsoft, are also building their own AI chips. On April 9th, Google also announced a new proprietary AI chip, which it said is currently available to developers and purpose-built for data centers.
ChatGPT’s maker, OpenAI, also has a plan to start producing chips for AI models. According to reports, OpenAI is looking to raise up to $7 trillion for the new AI chip business, which could become a big contender to Nvidia, the largest AI chip producer.
