NVIDIA NVLink Fusion Expands With NVHBM Custom High-Bandwidth Memory
nvidia ·
NVIDIA announced an expansion of its NVLink Fusion platform with NVHBM, a custom high-bandwidth memory technology designed for semi-custom AI infrastructure. Unlike traditional HBM, NVHBM integrates NVIDIA's custom memory controller into the 3D HBM base die rather than the XPU, freeing silicon area for compute. NVIDIA claims this approach delivers up to 30% greater memory bandwidth, 15% lower HBM power consumption, and up to 25% more XPU die area compared with standard HBM4E. A standard NVHBM implementation will be offered by multiple memory partners to reduce integration effort. Amazon's Annapurna Labs will be the first collaborator, supporting NVHBM and NVLink Fusion with its next-generation Trainium4 chips. This collaboration builds on AWS's previously announced support for NVLink Fusion, allowing Amazon chips and NVIDIA GPUs to share a common rack-scale architecture. NVLink Fusion gives partners access to NVIDIA chiplets, switches, MGX systems, and a broader ecosystem while they focus on custom XPU innovation.
nvidia