Nvidia Unveils NVHBM Custom Memory for NVLink Fusion Partners

Nvidia has introduced NVHBM, a custom high-bandwidth memory design aimed at partners building chips within its NVLink Fusion program. Rather than a replacement for standard high-bandwidth memory, the company positions NVHBM as an additional building block offered only to its custom silicon partners.

The core change is where the memory controller sits. In standard HBM, the controller occupies space on the compute die. According to TechPowerUp, NVHBM relocates that controller into the 3D HBM stack, which Nvidia says improves both performance and efficiency. Nvidia plans to use the technology in its own future GPUs and is extending it to third-party XPU customers through NVLink Fusion.

Compared with commodity HBM4e, Nvidia claims up to 30% more memory bandwidth per stack and 15% lower memory power consumption. TechPowerUp reports the company also cites up to 25% more usable compute die area on the processing unit. Those area savings come largely from a redesigned physical memory interface: Nvidia says the custom PHY cuts I/O area by up to 67% relative to JEDEC HBM4e, and that the narrower interface frees up to 80% more usable silicon across the layout by easing interposer routing.