Nvidia has introduced NVHBM, a toolkit for customized HBM4 memory that includes its own memory controller and an NVLink-Fusion interface, binding partners more tightly to Nvidia’s ecosystem.
Controller Embedded in the Base Die
HBM4 shares chip layers with the DRAM and uses a shared bottom logic layer for control. Unlike AMD’s approach, Nvidia embeds the memory controller directly into the base die. This frees space in CPUs, GPUs, and I/O dies. Miniaturized NVLink interfaces handle communication between the HBM and the compute dies.
All Three HBM Makers on Board
Separating the logic layer from the memory layers allows optimization with logic-focused manufacturing processes. With Samsung, SK Hynix, and Micron, all three HBM manufacturers are validating or offering NVHBM. Amazon’s semiconductor division Annapurna Labs is considered the first adopter.
A Strategy Against CXL
NVHBM is part of the NVLink-Fusion platform. With it, Nvidia binds companies designing their own AI accelerators to its interconnect technology and keeps NVLink prominent against competing standards such as Compute Express Link (CXL). Customers can customize but remain dependent on Nvidia’s interconnect technology.
Quellen: heise online.
