RealityHackerOpen in RealityHacker ⇢
Compute · AI Accelerators · published 2026-11-16T00:00:00+00:00 · via Nvidia Blog

NVIDIA Vera Rubin Achieves Record MLPerf Results in Training and Agentic Inference

Image via Nvidia Blog
Image via Nvidia Blog

The NVIDIA Vera Rubin platform delivered top performance across all MLPerf Training v6.1 benchmarks, outperforming the GB300 NVL72 system by up to 2x per GPU. In its first MLPerf Endpoints submission, Vera Rubin achieved over 10x improvement in throughput per watt for agentic AI workloads. The platform demonstrated strong scalability, maintaining 97% efficiency when scaling from 256 to 1,024 GPUs on the DeepSeek-V3 benchmark.

Expanded Detail

On November 16, 2026, NVIDIA unveiled its Vera Rubin platform in its first official benchmarking debut, achieving record-breaking results across all MLPerf Training v6.1 benchmarks and setting a new standard in agentic inference with MLPerf Endpoints. The Vera Rubin NVL72 system outperformed the prior-generation GB300 NVL72 by up to 2x per GPU in training workloads, demonstrating exceptional scalability—maintaining 97% efficiency when scaling from 256 to 1,024 GPUs on the DeepSeek-V3 model. In its inaugural Endpoints submission, Vera Rubin achieved over 10x improvement in throughput per watt for agentic AI tasks, such as those tested with the Kimi K3 model, while maintaining comparable response speeds (TTFT P90 of 4,673.9 ms) and user interactivity (50.4 tokens/sec per user). These results were verified by MLCommons Association for GB300 systems, while Vera Rubin’s data remains provisional pending peer review. The platform leverages NVIDIA’s CUDA software stack, NVLink scale-up networking, and Quantum-X800 InfiniBand/Spectrum-X Ethernet for large-scale coordination across thousands of GPUs—critical for maintaining efficiency at scale. Over 20 partners, including Microsoft Azure, Google Cloud, Dell, HPE, and Oracle, contributed to MLPerf Training submissions, underscoring the ecosystem’s reliance on NVIDIA’s open, programmable architecture. The broader context includes the growing demand for faster AI model training and efficient agentic workflows in sectors like drug discovery, enterprise IT, and customer service. While NVIDIA’s dominance is evident, the long-term impact hinges on whether these performance gains translate into real-world deployment efficiency and cost savings across diverse industries.

Context

The Vera Rubin platform’s performance leap could significantly accelerate the development and deployment of complex AI agents, potentially transforming how enterprises automate workflows, manage infrastructure, and deliver customer services. By drastically improving throughput per watt, NVIDIA may help reduce the environmental footprint of large-scale AI operations—a growing concern as data centers expand. The 10x efficiency gain in agentic inference could lower operational costs for cloud providers and enterprises alike, possibly enabling more widespread adoption of autonomous agents in real-time decision-making systems. This shift may also pressure competitors to innovate faster, potentially reshaping the semiconductor landscape beyond just GPU performance—toward integrated hardware-software optimization. A comparable precedent is the 2017 rise of the Tesla Autopilot chip, which demonstrated that specialized AI accelerators could outperform general-purpose chips in specific tasks, catalyzing a wave of custom silicon development. Here, Vera Rubin’s success may similarly encourage more investment in domain-specific AI hardware and software co-design. However, it remains uncertain whether these benchmarks reflect real-world latency or cost savings at scale, especially as power efficiency claims are still pending formal verification. What to watch for next is how quickly third-party developers and cloud providers integrate Vera Rubin into production systems—and whether the performance gains can be sustained without compromising model accuracy or reliability in mission-critical applications.

Expanded detail and Context are AI-generated analysis; the linked article remains the authoritative source.
Read the full article at Nvidia Blog →
Related stories
NVIDIA and Microsoft Extend AI Acceleration to Azure, Edge Devices, and Windows PCs · AI Accelerators
Lambda Enhances AI Infrastructure Resilience Using NVIDIA Photonics Switches · Networking & Interconnects
NVIDIA Unveils Optimized Deployment Tools for Top South Korean AI Models · AI Accelerators
This summary is Al-enhanced to contain extended analysis and broader social context. The original is {NAME); the linked article is the authoritative source. Original headline: “NVIDIA Sweeps MLPerf Training and Endpoints as Vera Rubin Debuts With Leading Performance.” Browse more stories.