Meta announces next-generation custom AI chips

Expanding its investments in artificial intelligence infrastructure, Meta today (10) Next-generation custom AI chips to optimize the language model training process. According to the statement, the updated version has “significantly better performance” than its predecessor.

The Meta Training and Inference Accelerator (MTIA) platform, introduced in May last year, Improve the computational efficiency of big tech’s AI systems. The processor also empowers developers to create models that result in new experiences across the company’s services.

With the new generation MTIA, the company train classification and recommendation models fasterincreases efficiency and output. The launch also achieves the desired balance between compute, memory bandwidth and memory bandwidth capacity.

Meta’s new MTIA chips, like their predecessors, will target data centers and complement the holding’s existing technological infrastructure. Also, they Designed to integrate with future developments in GPUsContributing to the next AI solutions from the owner of Facebook, Instagram, WhatsApp and Threads.

Three times more performance

The second-generation MTIA chip has 256 MB of on-chip memory running at 1.3 GHz, which is higher than the original version with 128 MB memory and 800 GHz. Three times more performance than with the first generationIn the first tests carried out by the technology giant.

The platform is currently used to train classification and recommendation algorithms, but this may change soon. The company plans to expand MTIA’s capabilities for use in productive AI solutionsWe are starting to embrace this in training llama models.

Rumors suggest that Meta is also working on the development of other AI chips, such as a project called Artemis that will be specific to inference missions. This market is currently dominated by NvidiaHowever, strong competitors such as Microsoft, Amazon and Google have invested heavily in this segment.