NVIDIA is reportedly revisiting the high‑bandwidth memory (HBM) configuration for its upcoming Rubin Ultra AI accelerators, scheduled for launch next year, as a response to tightening HBM supplies.
Rubin Ultra may adopt a different HBM setup
According to TrendForce, citing South Korean analysts, NVIDIA is evaluating several options to streamline the memory subsystem of Rubin Ultra.
- Switching from HBM4E to the baseline HBM4 standard.
- Using 8‑high (8Hi) HBM stacks instead of the originally planned 12‑high (12Hi) HBM4E.
No final decision has been made yet.
Reasons behind the potential shift
The primary driver is uncertainty surrounding the certification and mass production of HBM4E.
Leading DRAM manufacturers are still completing validation of the new memory, creating a risk that HBM4E shipments will not reach the volumes required for Rubin Ultra production ramp‑up.
Fewer layers mean more accelerators
Moving from 12‑high to 8‑high stacks reduces the number of DRAM dies per stack.
Consequently, a given quantity of HBM memory can yield more complete modules, increasing the potential output of AI accelerators.
Analysts note that NVIDIA already employed a similar tactic when scaling production of the LPDDR5X SOCAMM2 modules for its Vera CPU platform.
Bandwidth remains the priority
TrendForce argues that the chief advantage of Rubin Ultra will be its input‑output (I/O) speed rather than raw memory capacity.
Therefore, reducing the stack height could serve as a compromise that barely affects performance in most AI workloads while allowing a larger number of accelerators to reach the market.
Industry‑wide pressure
The report also highlights that other makers of specialized AI accelerators (ASICs) are examining comparable approaches amid the same HBM shortage.
Growing demand for AI infrastructure continues to strain supply chains, pushing vendors to balance peak performance with component availability.
Xpert Take
The tightening HBM supply is emerging as a critical bottleneck for the AI accelerator market. While earlier competition centered on GPU compute power, fast memory is now a decisive resource.
If TrendForce’s assessment proves accurate, NVIDIA’s move to a simpler HBM configuration would signal a pragmatic shift: favoring broader availability over a limited run of top‑spec parts.
For data‑center operators, the ability to deploy more accelerators—even with a modest reduction in per‑device specs—may outweigh the benefit of a few ultra‑high‑end models.
At present, the discussion remains speculative; NVIDIA has not officially confirmed any changes to Rubin Ultra’s specifications.
