How does the Rubin GPU architecture specifically accelerate agentic AI workloads beyond just raw compute power?

The Rubin GPU targets key bottlenecks in agentic workflows with several specialized features. It improves Mixture-of-Experts (MoE) scaling with an enhanced Tensor Memory Accelerator (TMA) that reduces data-movement overhead. For long-context tasks, it uses activation sparsity and adaptive compression to make attention mechanisms more efficient. Finally, it minimizes latency between computation steps using fine-grained dependent kernel triggering, which is critical for the sequential nature of AI reasoning tasks.

Inside NVIDIA Rubin GPU Architecture: Powering the Era of Agentic AI

NVIDIA Details Rubin GPU for Agentic AI Workloads

NVIDIA has detailed its next-generation Rubin GPU architecture, the core of the new Vera Rubin platform, designed to meet the growing demands of agentic AI. The company claims the new architecture will deliver up to a 10x improvement in agentic throughput per unit of energy compared to the current Blackwell platform. This development comes as the industry shifts from single-turn conversational AI to more complex, always-on AI factories that power multi-step reasoning, planning, and tool-use workflows, which demand sustained inference performance and efficiency.

Architectural Enhancements for Agentic Inference

The Rubin GPU is engineered to address critical bottlenecks in agentic inference, from data movement to long-context processing. Built with 336 billion transistors, the architecture integrates a new memory subsystem and connectivity technologies to keep its extensive compute resources fully utilized. Key architectural specifications include:

Compute: 224 streaming multiprocessors (SMs) and 896 Tensor Cores delivering up to 50 petaflops of NVFP4 performance.
Memory: Up to 288 GB of HBM4 memory providing 22 TB/s of peak bandwidth.
Connectivity: 3,600 GB/s of scale-up bandwidth via NVLink 6 and 1,800 GB/s for coherent CPU-GPU communication with NVLink-C2C.
Efficiency Features: Enhanced Tensor Memory Accelerator (TMA) with inline descriptor updates for MoE models, activation sparsity for long-context attention, and fine-grained kernel triggering to reduce latency.

System-Level Impact on AI Factories

The Vera Rubin platform extends these GPU-level advancements to the rack scale, treating the data center as a single, unified compute unit. The Vera Rubin NVL72 rack integrates liquid cooling and a cable-free MGX architecture to enable higher density, allowing for up to 40% more GPUs within the same power envelope. This system-level design is critical for deploying the multi-trillion parameter models that underpin sophisticated agentic systems, providing the resilient, high-throughput infrastructure required for continuous intelligence production.

With the Vera Rubin platform, NVIDIA is moving beyond discrete GPU performance to architecting the entire rack as a single, cohesive AI supercomputer, a necessary step to operationalize the high-throughput, always-on demands of agentic AI.

>> Verify Original Transmission at NVIDIA