Back to List
NVIDIA Blackwell Ultra NVL72 Sets Performance Record in Industry-First Agentic AI Benchmark AgentPerf
Industry NewsNVIDIABlackwellAgentic AI

NVIDIA Blackwell Ultra NVL72 Sets Performance Record in Industry-First Agentic AI Benchmark AgentPerf

NVIDIA has announced that its Blackwell Ultra NVL72 platform has secured a leading position in the inaugural AgentPerf benchmark, the industry's first standardized test for agentic AI infrastructure. Developed by Artificial Analysis, AgentPerf provides a comprehensive framework for developers, enterprises, and infrastructure providers to compare system performance across agentic AI workloads. In the first round of published results, the NVIDIA Blackwell Ultra NVL72 demonstrated exceptional efficiency, running 20x more agents per megawatt compared to previous NVIDIA systems. This benchmark marks a significant milestone in AI infrastructure evaluation, offering a clear metric for power efficiency and throughput as the industry shifts toward autonomous agentic applications.

NVIDIA Newsroom

Key Takeaways

  • First Industry Benchmark: AgentPerf, created by Artificial Analysis, is established as the first benchmark specifically designed to evaluate agentic AI infrastructure.
  • Blackwell Performance Leadership: The NVIDIA Blackwell Ultra NVL72 platform leads the initial round of testing, showcasing its dominance in agentic AI workloads.
  • Massive Efficiency Gains: The Blackwell platform achieves a 20x increase in the number of agents supported per megawatt compared to previous NVIDIA systems.
  • Strategic Utility: The benchmark provides a standardized way for enterprises and developers to compare and select AI infrastructure based on performance and efficiency.

In-Depth Analysis

The Emergence of AgentPerf as a Standard

As the AI landscape evolves from simple chatbots to complex, autonomous agents, the need for specialized benchmarking has become critical. AgentPerf, introduced by Artificial Analysis, fills this gap by becoming the industry’s first agentic AI benchmark. This tool is designed to provide a clear and objective way for developers, enterprises, and infrastructure providers to compare different systems. By focusing specifically on agentic AI workloads—which often require different computational profiles than standard LLM inference—AgentPerf allows stakeholders to make data-driven decisions about their hardware investments. The introduction of such a benchmark suggests a maturing industry where performance is no longer measured just by raw speed, but by the ability to handle the sophisticated logic and multi-step tasks inherent in agentic workflows.

Blackwell Ultra NVL72: Redefining Efficiency

In the first round of published results from the AgentPerf benchmark, the NVIDIA Blackwell Ultra NVL72 platform has emerged as the performance leader. A standout metric from the report is the platform's ability to run 20x more agents per megawatt than previous NVIDIA systems. This 20-fold increase in efficiency is a pivotal development for data center operators and enterprises concerned with the high energy demands of modern AI. The Blackwell Ultra NVL72 is engineered to maximize throughput while minimizing power consumption, a balance that is essential for scaling agentic AI applications. This performance lead indicates that the Blackwell architecture is specifically optimized for the high-concurrency and high-efficiency requirements of the next generation of AI agents.

Industry Impact

The results of the AgentPerf benchmark and the performance of the NVIDIA Blackwell platform have significant implications for the AI industry. First, the establishment of a standardized benchmark for agentic AI will likely accelerate the adoption of these technologies by providing enterprises with the confidence to evaluate and deploy infrastructure. Second, the 20x efficiency gain demonstrated by Blackwell sets a new bar for hardware providers, emphasizing that power efficiency is now as critical as computational power. As organizations look to deploy thousands or millions of autonomous agents, the ability to do so within reasonable power constraints will be the primary differentiator. This benchmark reinforces NVIDIA's position at the forefront of AI infrastructure, particularly as the market shifts toward more complex, agent-driven ecosystems.

Frequently Asked Questions

Question: What is AgentPerf and why is it important?

AgentPerf is the industry's first benchmark specifically designed for agentic AI infrastructure, developed by Artificial Analysis. It is important because it provides a standardized method for developers and enterprises to compare how different hardware systems handle the unique workloads associated with autonomous AI agents.

Question: How did the NVIDIA Blackwell Ultra NVL72 perform in the benchmark?

The NVIDIA Blackwell Ultra NVL72 platform delivered leading performance in the first round of AgentPerf results. Most notably, it demonstrated the ability to run 20x more agents per megawatt than previous NVIDIA systems, highlighting a massive leap in energy efficiency and throughput.

Question: Who can benefit from the AgentPerf benchmark results?

Developers, enterprises, and infrastructure providers can all benefit from these results. The benchmark offers a clear way to evaluate which systems are best suited for agentic AI workloads, helping organizations optimize their infrastructure for both performance and power consumption.

Related News

Meituan Unveils LongCat-2.0: A 1.6 Trillion Parameter Model Optimized for Agentic Coding on Domestic Clusters
Industry News

Meituan Unveils LongCat-2.0: A 1.6 Trillion Parameter Model Optimized for Agentic Coding on Domestic Clusters

Meituan's technology team has officially released LongCat-2.0, a landmark large language model featuring 1.6 trillion parameters. This model distinguishes itself as the first of its scale to complete the entire training and inference lifecycle on a domestic computing cluster of 50,000 cards. Designed specifically for Agentic Coding, LongCat-2.0 supports a native 1M long-context window and was pre-trained from scratch. With a dynamic activation range between 33B and 56B (averaging 48B), the model is engineered to provide high efficiency and stability in complex code understanding, generation, and execution tasks. This release marks a significant milestone for domestic AI infrastructure and the evolution of autonomous coding agents.

Meituan Technical Team Presents Selected Academic Papers at ICML 2026 to Advance Machine Learning Research
Industry News

Meituan Technical Team Presents Selected Academic Papers at ICML 2026 to Advance Machine Learning Research

The Meituan Technical Team has announced its participation in the International Conference on Machine Learning (ICML) 2026, one of the world's most influential academic gatherings in the field. ICML 2026 serves as a critical platform for discussing the future challenges and core issues facing machine learning development. Meituan's involvement includes the presentation of selected academic papers that have been evaluated for their significant theoretical value and practical impact. By contributing to this top-tier conference, the Meituan Technical Team aims to push the boundaries of the field and help lead future research directions. This engagement highlights the team's commitment to high-quality research that addresses both the fundamental questions of machine learning and its real-world applications, reinforcing their position within the global technical community.

Meituan Fulfillment AI Team Showcases LLM-Based Agent Innovations and Self-Evolving Systems at ACL 2026
Industry News

Meituan Fulfillment AI Team Showcases LLM-Based Agent Innovations and Self-Evolving Systems at ACL 2026

The Meituan Fulfillment AI Algorithm Team has unveiled its latest advancements in Large Language Model (LLM)-based Agent technology at a special session for the ACL 2026 conference. Focused on empowering Meituan's fulfillment business, the team is developing a self-evolving Agent operating system. Their research, which has resulted in dozens of publications in top-tier venues like ACL and EMNLP, spans critical domains including Continuous Pre-training (CPT), Post-training, Agentic Reinforcement Learning (RL), and Multimodal Understanding. This initiative represents a significant step in integrating frontier AI research with large-scale industrial fulfillment operations, aiming to enhance efficiency and system autonomy through advanced machine learning techniques.