NVIDIA Evaluates Downgrading HBM Configuration for Rubin Ultra AI GPUs
NVIDIA is reportedly evaluating a downgrade of the High Bandwidth Memory (HBM) specifications for its upcoming Rubin Ultra AI GPUs. The company is considering shifting from the originally planned 12-layer HBM4E to 8-layer HBM or standard HBM4 due to supply and verification challenges. This potential adjustment highlights the severe supply constraints and technical hurdles in the HBM supply chain, which could impact the performance and availability of next-generation AI hardware. By reducing the stack layers, NVIDIA can distribute the limited DRAM supply across more GPU units to meet high market demand. The downgrade options include transitioning from HBM4E to HBM4, or reducing the stack height from 12Hi to 8Hi HBM. Market research firm TrendForce suggests that since Rubin Ultra's primary upgrade lies in I/O speed, NVIDIA is more likely to opt for reducing the HBM stack layers.
## BACKGROUND
High Bandwidth Memory (HBM) is a 3D-stacked SDRAM architecture designed to provide ultra-fast data transfer rates for high-performance computing and AI workloads. The semiconductor industry has been facing a global HBM shortage driven by the rapid expansion of AI data centers. HBM4 and HBM4E represent the next generations of this technology, offering even higher bandwidth but presenting significant manufacturing and verification challenges for DRAM suppliers.