ASUS Announces NVIDIA HGX B300 AI Server Dominated MLPerf Training v6.0 Benchmarks with Top-Tier Performance

Delivering world-class performance for generative AI and LLM workloads

• World-leading AI performance: ASUS XA NB3I-E12 AI server achieved first-place rankings across key MLPerf Training v6.0 AI benchmarks
• Built for AI at scale: ASUS XA NB3I-E12 AI server with NVIDIA HGX B300 deliver exceptional training throughput and responsiveness
• Advancing AI innovation: ASUS drives transparent benchmarking and faster AI deployment through ongoing MLCommons participation

TAIPEI, Taiwan, June 22, 2026 — ASUS today announced that its high-performance XA NB3I-E12 AI server with NVIDIA HGX B300 has delivered outstanding results in the latest round of MLPerf Training v6.0 benchmarks. The system secured the No. 1 ranking across multiple critical categories, including Llama 3.1-8B (Server and Offline modes) and GPT-OSS-120B (Interactive mode). This incredibly powerful AI server is now shipping worldwide, further driving the continuous growth of ASUS as a world-class provider of AI solutions.

ASUS AI server sets new performance records across demanding AI workloads

Powered by eight NVIDIA Blackwell Ultra GPUs and dual Intel^® Xeon^® 6 Scalable processors, ASUS XA NB3I-E12 offers premium scalability and connectivity designed for demanding generative AI and LLM training. In its latest MLPerf submission, the system proved its architectural dominance by securing the No. 1 ranking across the industry’s most challenging scenarios:

• No. 1 performance, in Llama 3.1-8B (Server mode): Proves superior scheduling and high-concurrency handling in real-world, unpredictable multi-user environments.
• No. 1 throughput, in Llama 3.1-8B (Offline mode): Highlights the system’s ability to maximize GPU utilization and memory bandwidth under peak batch-processing workloads.
• No. 1 responsiveness, in GPT-OSS-120B (Interactive mode): Demonstrates zero-latency processing for complex reasoning in massive models, ensuring a seamless user experience.

In the latest submission, the system achieved optimized performance in 6.584 on the Llama 2 7B training task, a competitive 84.85 on the DeepSeek 671B ultra-large model test, and strong results of 74.75 on Llama 3.1 8B and 2.342 on the DLRM-DCN recommendation benchmark.

ASUS and MLPerf: Working in harmony to advance AI innovation

MLPerf is the industry-standard AI performance benchmark developed by MLCommons. As a long-term member, ASUS actively participates in the program to drive transparency and reproducibility across the industry. By contributing benchmark results and continuously optimizing its AI server platforms, ASUS remains committed to pushing the boundaries of AI technology, helping enterprises and researchers better evaluate infrastructure, and accelerating the deployment of responsible AI worldwide.