Jensen Huang Unveils Rubin: 5 Innovations, Performance Data, Agents & Robotics

Rubin platform overview

Rubin integrates six tightly coupled components—Vera CPU, Rubin GPU, NVLink 6 switch, ConnectX‑9 SuperNIC, BlueField‑4 DPU and Spectrum‑6 Ethernet—enabling “extreme codesign”. Compared with the Blackwell platform, Rubin reduces inference token cost up to ten‑fold and cuts the number of GPUs required for training mixture‑of‑experts (MoE) models by four‑fold.

Spectrum‑6 Ethernet uses 200 Gb/s SerDes, co‑packaged optics and an AI‑optimized architecture, delivering five‑fold energy‑efficiency, ten‑fold reliability improvement and five‑fold longer runtime.

NVLink 6 provides 3.6 TB/s per‑GPU bandwidth and 260 TB/s rack‑level bandwidth, with built‑in network compute for accelerated collective communication.

Vera CPU contains 88 custom Olympus cores, is Armv9.2‑compatible and supports NVLink‑C2C interconnect, offering industry‑leading CPU energy efficiency for AI workloads.

Rubin GPU incorporates the third‑generation Transformer Engine with adaptive compression, delivering 50 PFLOPS of NVFP4 compute performance.

The third‑generation confidential compute engine secures data across CPU, GPU and NVLink at rack scale.

The second‑generation RAS Engine provides real‑time health monitoring, fault tolerance and predictive maintenance for GPU, CPU and NVLink, and modular cable‑free trays accelerate system assembly up to 18× compared with Blackwell.

A new inference‑context memory storage platform driven by BlueField‑4 enables gigabyte‑scale KV‑Cache sharing and reuse across AI infrastructure, improving response latency and throughput with predictable low power consumption.

Open‑source ecosystem

Agentic AI – Nemotron series

Nemotron Speech : ASR model achieving ten‑fold speedup over comparable models on Daily and Modal benchmarks.

Nemotron RAG : New embedding model and re‑ranking visual‑language model for multilingual, multimodal retrieval, improving document search precision.

Nemotron Safety : Llama‑Nemotron content‑safety model supporting multiple languages and Nemotron PII for high‑precision sensitive‑data detection.

Physical AI & robotics – Cosmos series

Cosmos Reason 2 : Leading inference VLM for robot perception, enhancing accuracy of perception, understanding and interaction.

Cosmos Transfer 2.5 and Predict 2.5 : Large‑scale synthetic video generation for diverse environments.

Isaac GR00T N1.6 : Open‑source VLA model for humanoid robots that provides full‑body control and leverages Cosmos Reason for context understanding.

NVIDIA Blueprint : Video search and summarization workflow within Metropolis platform for analyzing recorded and live video streams.

Autonomous driving – Alpamayo series

Alpamayo 1 : Open‑source VLA model for autonomous vehicles that can both perceive the environment and explain its own actions.

AlpaSim : Closed‑loop simulation framework supporting training and evaluation of inference‑type driving models in diverse edge scenarios.

Physical‑AI dataset of over 1,700 hours of real‑world driving data covering rare and complex conditions.

AI for Science – Clara AI suite

La‑Proteina : Atomic‑scale protein design model for drug candidate discovery.

ReaSyn v2 : Incorporates manufacturability constraints to ensure AI‑designed molecules are synthesizable.

KERMT : Predicts drug‑human interaction for early‑stage safety testing.

RNAPro : Predicts RNA three‑dimensional structures to enable personalized medicine.

Synthetic protein structure dataset containing 455 k entries to support model training.

Implications

The Rubin platform demonstrates a shift from pure model scaling toward efficient, stable AI infrastructure. By organizing compute, compressing costs, moving models toward inference‑centric deployment, and deepening the coupling between agents and the physical world, Rubin aims to enable large‑scale AI applications with lower operational overhead.

References

https://nvidianews.nvidia.com/news/rubin-platform-ai-supercomputer

https://blogs.nvidia.com/blog/open-models-data-tools-accelerate-ai/