Nvidia will abandon HBM4 speed for sufficient memory volume for Vera Rubin

New data from ZDNet in South Korea shows how Nvidia is changing its HBM4 memory requirements for the Vera Rubin accelerator.

What was originally planned?

The company intended to supply accelerators with HBM4 memory at a transfer rate of no less than 11.7 Gb/s. Suppliers approached were SK Hynix, Samsung Electronics and Micron Technology—all capable of producing such chips.

Problem with volumes

However none of the three manufacturers can yet deliver the required quantities of memory at the required speed. Therefore Nvidia is deciding to soften its requirements.

New conditions

Now the company is willing to accept HBM4 running only at 10.6 Gb/s, while remaining open to faster chips (11.7 Gb/s). This was done in response to an increased memory shortage this year: Nvidia does not want to complicate deliveries for itself, its partners and customers with overly strict specifications.

Supplier capabilities

- SK Hynix can provide up to 60 % of the needed HBM4 volumes, but not in the fully fast variant.
- Samsung announced the start of serial production, but accelerating capacity growth in the short term is difficult for it.
- Micron remains a less reliable source under such high requirements.

Thus Nvidia is forced to adapt to the real capabilities of HBM4 manufacturers, which may affect the operating speed of some Vera Rubin accelerators.