The Nvidia logo. [AFP]
The Nvidia logo. [AFP]

Nvidia is considering equipping its next-generation GPU, the Rubin Ultra, with less memory than originally planned as it braces for a potential shortage of high-bandwidth memory chips, The Information reported Thursday.

Three sources told the outlet that Nvidia has been testing at least three Rubin Ultra GPU samples in recent weeks, and that some of those samples carry less memory than the company announced when it first unveiled the chip.

One reason Nvidia is weighing a reduced memory specification, the sources said, is that memory suppliers may struggle to produce enough HBM4E in time for the Rubin Ultra's scheduled launch. While lower memory specs could affect performance, Nvidia's customers noted that the trade-off could benefit buyers looking to cut costs, as a less memory-intensive chip would likely carry a lower price tag.

Nvidia CEO Jensen Huang first unveiled the Rubin Ultra at a developer conference last year, saying each GPU would be equipped with 1 terabyte (TB) of HBM4E. The chip was originally designed to include 16 memory stacks.

The samples currently under testing, however, feature fewer stacks and less memory capacity per die. Some samples use HBM4 rather than HBM4E, the sources said.

According to the sources and one Nvidia customer, total memory capacity on some tested samples reached only 192 gigabytes (GB), while others came in at around 256 GB — well below the originally announced 1 TB.

Nvidia's current mass-production chip, the Vera Rubin, supports up to 288 GB of HBM4, meaning the Rubin Ultra test versions would actually carry less memory than their predecessor. The figures represent a reduction of 75 to 81 percent from the original plan and 11 to 33 percent compared with the previous generation. That stands in contrast to remarks by Andrew Bell, Nvidia's senior vice president of hardware engineering, who said in mid-July that the company had been "getting ahead of the memory problem" and saw no supply-side issues "for now."

HBM4E is designed to move more data and consume less power than standard HBM4, but achieving that requires denser chip construction and faster electrical connections, making packaging significantly more complex. Technology research firm Epoch AI has assessed that HBM can account for more than half the component cost of advanced AI chips, and that a lower-memory version could still suit a wide range of AI applications while substantially reducing the price.

To address the supply constraints, Nvidia struck a $500 billion partnership with SK Group late last month to jointly develop HBM4 and HBM4E in multiple configurations.

Raj Mirpuri, Nvidia's vice president of global AI cloud and infrastructure, said the partnership was aimed at "securing a stable supply of HBM" and that SK Hynix would invest in expanding its HBM production capacity for Nvidia.

SK Hynix announced in June that it plans to double its memory chip production capacity within five years.

Nvidia has not yet finalized the Rubin Ultra's specifications or pricing, and the sources said the company still has time to adjust the design based on memory supply conditions, costs and customer demand before shipments begin late next year.


yckim6452@heraldcorp.com