NVIDIA NVHBM Targets 30% More Bandwidth Than HBM4E

Posted on
NVIDIA introduces NVHBM, a custom memory design promising 30% more bandwidth, 15% lower power, and more usable compute die area.

Quick Report

NVIDIA has introduced NVHBM, a custom high-bandwidth memory design that moves the memory controller into the HBM stack instead of placing it on the compute die. NVIDIA claims up to 30% more bandwidth, 15% lower HBM power consumption, and up to 25% more usable compute die area compared with standard HBM4E.

The redesigned physical interface reportedly reduces I/O area by up to 67% and simplifies interposer routing. NVIDIA says NVHBM could reach roughly 5.2 TB/s per stack when paired with 16 Gbps HBM4E, while the technology will also be offered to third-party XPU customers through NVLink Fusion, with Amazon's Annapurna Labs named as an initial partner.

Written using GitHub Copilot GPT-5 mini in agentic mode instructed to follow current codebase style and conventions for writing articles.

Source(s)

  • TPU