Nvidia Unveils NVHBM as AWS Trainium4 Becomes First Adopter

Demand for high-bandwidth memory is rising as generative AI workloads require faster data movement, while cloud providers are developing custom accelerators to reduce reliance on general-purpose GPUs. Nvidia’s NVHBM integrates the memory controller into the HBM logic die, a design intended to increase bandwidth and free up more area on the compute die. The approach extends Nvidia’s technology into chips designed by customers that also compete with its core accelerator business.

Amazon’s chip-design unit Annapurna Labs will be the first partner to adopt NVHBM, with plans to use the technology in AWS’s next-generation Trainium4 AI training chip. Nvidia did not disclose a shipment date, pricing or the financial terms of the arrangement. The partnership shows how Nvidia is seeking a role in cloud companies’ in-house silicon programs through memory and system integration, even as those customers expand alternatives to Nvidia GPUs.