Meta to use custom AMD Instinct MI400 accelerators with 144GB HBM4 for select workloads, report claims

According to a report by SemiAnalysis, Meta will use a custom version of AMD's Instinct MI450-series accelerators (likely MI455X) with only 144GB of HBM4 memory, one-third of the standard 432GB configuration. The custom chip is optimized for Facebook's recommendation systems, offering better bandwidth-per-dollar and lower power consumption, but with significantly reduced compute and memory capacity. This specialization makes it unsuitable for training frontier AI models or running LLM inference. While the custom design could save Meta tens of millions of dollars in bill-of-materials costs, it reduces versatility and interchangeability. As a result, Meta may still need to purchase general-purpose accelerators—potentially from Nvidia—for its frontier AI workloads, creating an ironic situation where cost-cutting in one area drives spending in another.