Rubin Ultra Makes Major Cuts, Even Nvidia Can't Handle Memory Price Hikes?
NVIDIA's Rubin Ultra, the top-tier variant of the newly announced Rubin AI accelerators, has reportedly seen significant specification downgrades, according to an industry report from SemiAnalysis. Initially designed with four compute dies (4-die), the Rubin Ultra is now said to be reduced to a 2-die design.
Key changes highlighted in the report include:
* **No increase in peak theoretical compute performance**, remaining at 35 PFLOPs like the standard Rubin.
* **Severe reduction in memory capacity** to 192GB using 8-Hi HBM stacks, which is less than the standard Rubin's 288GB using 12-Hi stacks.
* **Negligible memory bandwidth improvement** of only 1 TB/s.
* **Slightly higher chip-level power consumption**.
* The **primary upgrade is a massive increase in scale-up interconnect capacity**, supporting connections for up to 576 GPUs via NVLink, compared to 72 for the standard Rubin.
The report suggests the redesign is primarily a cost-optimization move driven by the sharp rise in HBM (High-Bandwidth Memory) prices. By reducing the expensive HBM content and shifting investment towards enhanced system-scale networking, NVIDIA aims to maintain the platform's value for large-scale AI training clusters while managing soaring material costs.
The news reportedly triggered a sell-off in South Korean memory stocks, with SK Hynix and Samsung shares falling around 8%, as markets grew concerned that NVIDIA—a major HBM buyer—might be reducing its reliance on high-capacity memory, potentially capping future pricing power for memory makers.
Odaily星球日报08/03 05:01
