NVIDIA Cuts HBM and NVLink Bandwidth for Rubin GPU Systems, Boosts Network Bandwidth
NVIDIA's Rubin GPU system HBM4 memory bandwidth has been reduced from 44TB/s to 38.5TB/s, NVLink bandwidth has dropped from 7.2TB/s to 6TB/s, while scale-out network bandwidth has increased from 0.8TB/s to 0.9TB/s.
Key Takeaways
- Key Highlight：NVIDIA's Rubin GPU system HBM4 memory bandwidth has been reduced from 44TB/s to 38.5TB/s, NVLink bandwidth has dropped from 7.2TB/s to 6TB/s, while scale-out network bandwidth has increased from 0.8TB/s to 0.9TB/s.
- Innovation & Tech：Highlights advancements in NVIDIA, Cuts, HBM, demonstrating rapid progress in model capabilities.
- Industry Impact：Reported via IT之家 (CN), offering actionable signals for developers and technology leaders.
Specifications for NVIDIA's next-generation Rubin GPU architecture have been adjusted. A comparison between official documentation from October this year and third-party archived documents from June shows that the Vera Rubin superchip's HBM4 memory bandwidth has been reduced from 44TB/s to 38.5TB/s, NVLink interconnect bandwidth has dropped from 7.2TB/s to 6TB/s, while scale-out network bandwidth has increased from 0.8TB/s to 0.9TB/s. These changes are reflected proportionally across both the NVL4 and NVL72 configurations.
The HBM bandwidth reduction is approximately 12.5%, and the NVLink bandwidth reduction is approximately 16.7%, which represents a relatively significant specification regression in the AI training chip space. Possible reasons include HBM4 mass production yields not yet meeting expectations, packaging process challenges, or NVIDIA rebalancing design targets across subsystems under power and thermal constraints. The increase in network bandwidth indicates that NVIDIA continues to invest in cluster interconnect capabilities.
For AI large model training, HBM bandwidth directly affects data throughput during model inference and training, and the bandwidth reduction may mean per-GPU performance falls short of earlier expectations. However, the trade-off between NVLink and scale-out network bandwidth suggests that NVIDIA may be tilting more design resources toward cluster-level interconnect, aiming to compensate for reduced per-GPU bandwidth through multi-GPU collaboration.
The Rubin architecture is NVIDIA's next-generation GPU platform following Hopper and Blackwell, and its specification adjustments have implications for the AI compute supply chain and data center construction planning. Downstream customers evaluating procurement timelines and architecture compatibility should pay attention to the actual parameters of the final mass-production version.
This page provides an editorial summary based on publicly available information. It is not a republished article. Use the source link below for the original report.
Industry Insights & Analysis
As artificial intelligence rapidly evolves, breakthroughs surrounding NVIDIA, Cuts, HBM, NVLink are shifting toward scalable, robust real-world implementations.
Driven by both open-source ecosystems and proprietary model architectures, the integration between compute optimization, data engineering, and agentic workflows is accelerating. This development provides a strategic benchmark for upcoming AI tooling and developer workflows.