NVIDIA Launches NVHBM, AWS to Deploy 2 Million More GPUs

According to NVIDIA, the company unveiled its next-generation custom high-bandwidth memory (HBM) technology, NVHBM, on August 26 and integrated it into the NVLink Fusion ecosystem. Unlike conventional HBM architectures, NVHBM moves the custom memory controller from the XPU compute die to the base die within the HBM stack.

NVHBM Reworks HBM Architecture

In traditional HBM designs, the memory controller is located on the XPU compute die, consuming valuable silicon area that could otherwise be used for additional computing resources. NVIDIA said NVHBM addresses this limitation by relocating the controller to the HBM base die.

Compared with standard HBM4E, NVHBM can deliver up to 30% higher memory bandwidth, 15% lower HBM power consumption, and free up to 25% more XPU die area for compute units.

NVIDIA also developed a customized physical layer (PHY) interface that reduces I/O area by 67% compared with the JEDEC-standard HBM4E interface. The design can simplify interconnect routing in advanced packaging and further improve silicon-area utilization.

NVIDIA said NVHBM has been validated by multiple major memory manufacturers and is being developed on a technology foundation aligned with its future GPU platforms.

AWS to Deploy 2 Million Additional NVIDIA GPUs

Amazon's chip design unit, Annapurna Labs, is the first partner to adopt NVHBM. Starting with its next-generation Trainium4 platform, Annapurna Labs will support NVLink Fusion, enabling Amazon's custom AI chips to work alongside NVIDIA GPUs within a rack-scale architecture.

Meanwhile, AWS and NVIDIA announced plans to deploy an additional 2 million NVIDIA GPUs between 2027 and 2028 across AWS infrastructure. The deployment will cover NVIDIA's Blackwell Ultra, Rubin, and Rubin Ultra architectures.

The two companies will also expand cooperation in AI factories, CPUs, networking, open models, data processing, and robotics. AWS plans to combine NVIDIA GPUs and its own Trainium processors, giving customers greater flexibility to select computing architectures based on specific AI workloads.

NVIDIA Expands Beyond GPUs

NVHBM represents another step in NVIDIA's strategy to shape the architecture of customized AI systems. Through NVLink Fusion, partners can integrate NVIDIA technologies such as NVLink Chiplet, NVLink-C2C, NVLink Switch, and MGX rack-scale systems.

NVHBM adds memory architecture to this ecosystem, extending NVIDIA's influence from individual GPUs toward system-level AI infrastructure.

The technology is not intended to replace standard HBM products. Instead, it is designed as a customized memory solution for future AI chip platforms. NVIDIA's current Vera Rubin rack-scale systems do not use NVHBM, indicating that the technology is primarily aimed at future custom chip designs.