What happened
Nvidia is reportedly testing Rubin Ultra chip designs with lower memory capacities. Some tested options drop memory to 192 GB. Other designs step back to standard HBM4 memory specs. High-bandwidth memory shortages are driving these design adjustments.
The context
High-bandwidth memory remains a tight bottleneck across the AI hardware industry. Trimming memory specs per chip can help Nvidia ship more units, but it lowers the available memory bandwidth for running massive AI models.
Sources
- Tom's Hardware ↗ via Google News Reported