Nvidia Reportedly Tests Lower HBM Configurations for Rubin Ultra
TL;DR. Nvidia is reportedly testing versions of its upcoming Rubin Ultra AI accelerator with reduced HBM memory due to anticipated supply shortages. - Some designs feature as little as 192 GB of memory, a decrease from the previously announced 1 TB. - Nvidia may also use HBM4 instead of the more advanced HBM4E to mitigate supply constraints. - The memory reduction could impact the performance and capabilities of the Rubin Ultra accelerator in AI data centers.
- Nvidia is reportedly evaluating Rubin Ultra configurations with less HBM due to supply concerns.
- Tested designs include 192 GB or 256 GB of memory, down from the original 1 TB.
- The company may also revert to HBM4 from HBM4E for some versions.
- Memory shortages for high-bandwidth memory are affecting AI accelerator development.