The NVIDIA Ruby platform
The new standard for agentic AI and logical reasoning
In the world of high-performance computing (HPC), standing still is synonymous with falling behind. While the Blackwell architecture is just beginning to conquer data centres, NVIDIA has already announced the next big leap: the NVIDIA Rubin platform.
At sysGen, we analyse why Rubin is not just a hardware upgrade, but the foundation for the next stage of AI – known as agentic AI.
comming soon
Source and further information:
NVIDIA Rubin Plattform
For more details, please see the official announcement video:
Announcement video
With the Rubin platform, NVIDIA no longer considers the individual chip as the benchmark, but rather the entire data centre. This system-centric approach makes it possible to efficiently handle complex, multi-stage problem-solving processes and workflows with extremely long contexts.
An overview of technological breakthroughs:
The core of the platform utilises the new HBM4 memory. This eliminates the critical bottleneck of memory bandwidth, resulting in massively accelerated inference – more tokens per watt with decreasing cost per token.
The new Vera CPU has 88 NVIDIA-designed cores and offers a memory bandwidth of up to 1.2 TB/s (LPDDR5X). Thanks to NVLink-C2C connectivity, it works perfectly with Rubin GPUs.
The new interconnect doubles the performance compared to Blackwell. It offers a bandwidth of 3.6 TB/s per GPU. In an NVL72 system, 72 GPUs are combined into a single, gigantic performance domain.
With hardware-accelerated adaptive compression, it enables NVFP4 inference of up to 50 petaFLOPS – while remaining fully compatible with existing Blackwell optimisations.
Proactive real-time maintenance ensures maximum reliability. The new cable-free tray design in the rack enables up to 18 times faster installation and maintenance.
NVIDIA offers Rubin technology in various form factors to meet different scalability and performance requirements:
- NVIDIA Vera Rubin NVL72: The rack-scale solution that connects 72 GPUs and 36 Vera CPUs – ideal for industrial AI in the gigascale range.
- NVIDIA DGX Vera Rubin NVL72: The turnkey infrastructure solution for companies that want to train and deploy complex models as quickly as possible.
- NVIDIA DGX Rubin NVL8: A liquid-cooled system with eight Rubin GPUs, optimized for training and inference in a more compact format.
The Rubin platform is designed to dramatically reduce the cost per token while increasing the intelligence of the models through efficient reasoning. For our customers, this means:
- Maximum efficiency: More computing power with lower energy consumption per operation.
- Future-proofing: Seamless transition from Blackwell to Rubin workloads.
- Security: Third-generation confidential computing protects your proprietary data across the entire rack (CPU, GPU and NVLink).
With Rubin, NVIDIA provides the necessary infrastructure for AI systems that not only respond, but actively solve problems. At sysGen, we support you in harnessing this enormous power for your specific requirements.