Meta Reportedly Customizes AMD MI450 GPU for Recommendation Systems
Meta is reportedly customizing AMD's upcoming MI450-architecture GPU, reducing it to half the compute units and one-third the memory capacity of the standard MI455. This custom design features four compute dies and six HBM4 memory stacks instead of the standard eight dies and twelve stacks. This customization highlights a growing trend of hyperscalers tailoring hardware for specific workloads, optimizing for recommendation systems (RecSys) rather than large language models (LLMs). By adjusting the CPU-to-GPU compute ratio and memory bandwidth, Meta aims to improve cost-efficiency for its massive recommendation workloads. The standard Instinct MI455 features eight TSMC N2 compute dies and 432GB of HBM4 memory, whereas Meta's variant scales down to four compute dies and 144GB of HBM4 memory. While this configuration improves efficiency for recommendation algorithms, it is less suitable for LLM training and inference.
## BACKGROUND
Recommendation systems (RecSys) and Large Language Models (LLMs) have fundamentally different hardware requirements. While LLMs demand massive compute power and memory capacity to process dense transformer models, recommendation systems often rely on large, sparse embedding tables that require high memory bandwidth and balanced CPU-to-GPU communication rather than raw GPU compute. AMD's Instinct MI400 series, utilizing next-generation HBM4 memory, serves as the foundation for these high-performance AI workloads.