Rambus Inc. has launched its HBM4E Memory Controller IP, a new technology designed to tackle critical memory bandwidth challenges. This innovation aims to boost performance and reliability for demanding AI accelerators, GPUs, and high-performance computing systems.
The controller supports speeds up to 16 gigabits per second per pin, enabling memory bandwidth of over 4.1 terabytes per second per HBM4E device. In an AI accelerator setup with eight HBM4E stacks, this configuration can achieve more than 32 terabytes per second, crucial for large-scale AI model training and data-intensive computing.
High-bandwidth memory is fundamental to modern AI silicon, bridging the gap between processing power and memory access speeds. Rambus's controller IP is positioned to support future AI system-on-chip designs where memory performance is a key differentiator.
Building on over 100 HBM design wins, the new controller integrates advanced reliability features to ensure first-time silicon success, a vital aspect given rising chip complexity and manufacturing costs. The IP is available for licensing immediately.