NVIDIA has unveiled NVHBM, a custom High Bandwidth Memory solution developed alongside Amazon's Annapurna Labs. An extension of NVLink Fusion, this technology integrates the memory controller into the HBM base die, removing it from the main XPU architecture.
The NVHBM implementation offers significant efficiency gains, including 30% more bandwidth and 15% lower power consumption than standard HBM4E. By relocating the controller, NVIDIA saves 25% die area on the main chip for increased processing density and scale.
Amazon's Trainium4 chips will be among the first to utilize NVHBM via NVLink Fusion. NVIDIA plans to standardize this for multiple providers, with the 2028 Feynman GPU generation expected to adopt the technology to optimize future AI infrastructure performance.