Nvidia unveils custom NVHBM memory with 30% more bandwidth and 15% less power

How NVHBM Delivers Performance Gains

Nvidia has introduced NVHBM, a custom high-bandwidth memory solution designed to outperform standard HBM4e chips. The technology was revealed as part of the company's NVLink Fusion partner program, targeting data center and AI workloads. Unlike traditional commodity memory, NVHBM uses a custom base die and physical interface to deliver improved performance and efficiency.

The new memory architecture promises up to 30% higher bandwidth compared to off-the-shelf HBM4e components while consuming 15% less power during operation. These gains stem from Nvidia's decision to build a specialized base die and PHY layer tailored specifically for its compute platforms. The design is being made available to partners within the NVLink Fusion ecosystem, allowing them to integrate NVHBM into their own systems and accelerators.

NVHBM's improvements come from rearchitecting the memory subsystem at the physical level. By replacing standard base die components with custom silicon, Nvidia can optimize signal paths and reduce latency between memory and compute units. The custom PHY layer also enables tighter integration with NVLink interconnects, which are critical for scaling AI training across multiple GPUs. This approach allows for more efficient data movement and lower energy per bit transferred.

Will NVHBM Replace Standard HBM?

While NVHBM offers compelling advantages, Nvidia emphasizes that it is not intended as a direct replacement for commodity HBM4e. Instead, the custom memory solution is positioned as an optional upgrade path for partners building high-performance systems. Standard HBM will remain available for broader market applications where cost and compatibility take priority over peak performance.

The introduction of NVHBM reflects Nvidia's ongoing strategy to vertically integrate more of its hardware stack, giving it greater control over performance and power efficiency. As AI workloads grow more demanding, custom memory solutions like NVHBM may become essential for maintaining competitive advantage.

Frequently Asked Questions

What is NVHBM and how does it differ from standard HBM4e? NVHBM is Nvidia's custom high-bandwidth memory implementation that uses a specialized base die and PHY to achieve 30% more bandwidth and 15% lower power than commodity HBM4e. It is not a drop-in replacement but an enhanced option for select partners.

Which partners can access NVHBM? NVHBM will be available to members of Nvidia's NVLink Fusion partner program, enabling them to incorporate the custom memory into their own data center and AI hardware designs.

When will systems using NVHBM be available? Nvidia has not announced specific timelines for NVHBM-equipped systems, but the technology is being rolled out alongside the NVLink Fusion initiative targeting advanced computing deployments.