Nvidia Tests Lower Memory Specs for Rubin Ultra Amid HBM Shortage

Nvidia is reportedly testing significantly lower memory configurations for its upcoming Rubin Ultra GPU to combat a critical shortage of high-bandwidth memory (HBM). The move suggests that supply chain constraints are forcing the chipmaker to reconsider the specifications of its next-generation AI hardware.

According to industry reports, including those from Tom's Hardware, Nvidia is evaluating Rubin Ultra designs with memory capacities ranging from 192 GB to 256 GB. This represents a steep decline from the original specifications, which outlined a massive 1 TB of HBM4e memory. That initial design was intended to utilize 16 stacks of 16-Hi memory, with 64 GB per stack and eight stacks per reticle die. To further ensure supply stability, some of the versions currently being tested may step back from HBM4e to standard HBM4.

The HBM Bottleneck

The Rubin architecture is the successor to the Blackwell platform, with the "Ultra" variant designed as a high-end fusion of GPUs tailored for the most demanding AI workloads. However, the industry is currently grappling with a severe HBM shortage. Major memory suppliers, including SK Hynix, Samsung, and Micron, are reportedly sold out of capacity through 2027, leaving chip designers struggling to secure the specialized memory required for top-tier accelerators.

Implications for AI Scaling

A reduction in memory from 1 TB to as little as 192 GB could have significant consequences for the deployment of the largest large language models (LLMs). Memory capacity is critical for holding model weights; a drastic cut may force developers to rely more heavily on slower system memory or implement more complex model sharding across multiple GPUs to maintain performance. This shift highlights a growing trend in the AI hardware market where memory availability, rather than raw compute power or FLOPS, has become the primary bottleneck for next-generation accelerators.

What to Watch

While Nvidia may frame these adjustments as the creation of different SKUs or "optimizations" for various customer needs, the gap between the original 1 TB target and the 192 GB tests is substantial. It remains to be seen whether these lower-capacity versions will become the standard for the Rubin Ultra or if Nvidia can secure enough HBM4e to deliver the original specifications. Industry observers will be watching for official SKU announcements to determine how much of the original vision for the Rubin Ultra will survive the current supply crisis.