Meta's custom AMD GPU cuts compute and memory, focuses on recommendation systems

Research firm SemiAnalysis says Meta's custom AMD Instinct GPU, based on the MI450 architecture, is very different from the standard MI455. The standard version has 8 compute chips and 432GB of HBM4 memory, but Meta's custom version only has 4 chips and 144GB memory, cutting compute power in half and memory by two-thirds.

The design aims to balance CPU and GPU power in rack systems, making memory bandwidth more cost-effective for Meta's recommendation systems (RecSys). However, SemiAnalysis notes this setup isn't great for training or running large language models (LLMs).