Rambus has unveiled its new HBM4E Memory Controller IP, claiming industry leadership in high-bandwidth memory technology for next-generation AI accelerators and GPUs. The controller supports speeds up to 16 Gbps per pin, delivering unprecedented memory bandwidth for demanding AI workloads. This advancement addresses the critical memory bottleneck that limits large language model performance.

Industry leaders are already praising the development. Samsung's Foundry IP Development Team emphasized that HBM4E IP solutions will be essential for broad industry adoption, while MatX CEO Reiner Pope noted that memory bandwidth is one of the main bottlenecks on LLM performance. IDC's Soo Kyoum Kim highlighted that as AI processor requirements rapidly increase, HBM solutions must advance apace. The Rambus HBM4E Controller can achieve 4.1 Terabytes per second throughput per memory device.

For AI accelerators using eight HBM4E devices, this translates to over 32 TB/s of memory bandwidth for next-generation AI workloads. The controller IP can be paired with third-party PHY solutions to create complete HBM4E memory subsystems in 2.5D or 3D packages. Rambus is making the HBM4E Controller available for licensing immediately, with early access design customers able to engage now.