Nvidia is exploring options to reduce the memory capacity of its next-generation artificial intelligence (AI) chip, 'Rubin Ultra,' in response to a shortage of high-bandwidth memory (HBM).
On August 6, U.S. technology news outlet The Information reported, citing three sources, that Nvidia has tested at least three prototypes of Rubin Ultra with varying memory capacities in recent weeks.
Some prototypes are equipped with 192 gigabytes (GB) of HBM per graphics processing unit (GPU), while others feature 256GB. It is also reported that some prototypes are using HBM4, an earlier version, instead of the originally planned next-generation HBM4E.
Nvidia's testing of multiple specifications is a precaution against the possibility of not securing enough HBM4E by the time Rubin Ultra is released. The demand for HBM in AI semiconductors has surged, impacting the supply for next-generation GPU designs.
Last year, Nvidia announced plans to equip Rubin Ultra with approximately 256GB of HBM4E per GPU, totaling up to 1 terabyte (TB). The 192GB product currently being tested represents a 25% reduction in memory capacity from the original plan.
Considering that the currently produced 'Vera Rubin' can accommodate up to 288GB of HBM4 per GPU, the memory capacity of the Rubin Ultra prototypes is 11% to 33% lower than existing products.
A reduction in memory capacity could lead to decreased performance when processing large AI models, but it may also lower product prices. Some Nvidia customers have reportedly expressed that lower-capacity products could help reduce costs.
EpochAI, an AI technology research organization, has analyzed that HBM can account for more than half of the manufacturing costs of advanced AI chips. Thus, the supply and pricing of HBM have become critical factors influencing both the performance and sales prices of AI chips.
Nvidia is also working to secure HBM supplies. Last month, the company announced a large-scale AI collaboration plan with SK Group and is pursuing the development and expansion of next-generation memory, including HBM4 and HBM4E, in partnership with SK Hynix. SK Hynix also plans to double its memory production capacity over the next five years.
The final memory capacity and pricing for Rubin Ultra have yet to be determined. Nvidia is expected to adjust specifications based on the supply situation, costs, and customer demand for HBM4E before shipments in the second half of next year.
* This article has been translated by AI.
Copyright ⓒ Aju Press All rights reserved.