ASUS XA NB3I-E12 AI Server Powered by NVIDIA HGX B300 Tops MLPerf Training v6.0 Across Llama 3.1-8B and GPT-OSS-120B

TLDR

• ASUS XA NB3I-E12 with NVIDIA HGX B300 secured #1 rankings in three MLPerf Training v6.0 categories: Llama 3.1-8B Server, Llama 3.1-8B Offline, and GPT-OSS-120B Interactive
• Powered by eight NVIDIA Blackwell Ultra GPUs and dual Intel Xeon 6 Scalable processors, targeting large-scale generative AI and LLM training
• Submission also logged strong scores on Llama 2 7B (6.584), DeepSeek 671B (84.85), Llama 3.1 8B (74.75), and DLRM-DCN (2.342)
• ASUS XA NB3I-E12 is shipping worldwide now; enterprise pricing is via local ASUS representatives

ASUS Claims Three #1 Spots in MLPerf Training v6.0

ASUS has landed at the top of the MLPerf Training v6.0 leaderboards with its high-performance XA NB3I-E12 AI server, the company’s flagship rackmount system built around NVIDIA’s HGX B300 platform. The system secured the No. 1 ranking across three of the benchmark suite’s most challenging categories: Llama 3.1-8B in Server mode (real-world high-concurrency scheduling), Llama 3.1-8B in Offline mode (peak batch-processing throughput), and GPT-OSS-120B in Interactive mode (zero-latency reasoning for massive models). For enterprises evaluating AI infrastructure, that combination of wins covers most of the deployment scenarios that matter: multi-user serving, batch inference, and interactive reasoning on the largest open-weight models.

The XA NB3I-E12 is built on eight NVIDIA Blackwell Ultra GPUs paired with dual Intel Xeon 6 Scalable processors, configured for premium scalability and connectivity designed specifically for demanding generative AI and LLM training workloads. The architecture is the kind that hyperscalers, sovereign AI clouds, and large enterprise research teams deploy when they need to push model training throughput to the limit. The #1 ranking in the Server mode category is particularly meaningful because it reflects performance under unpredictable multi-user load, not just synthetic peak throughput.

Benchmark Numbers Worth a Closer Look

The submission logged competitive results across four additional benchmarks beyond the three category wins. On the Llama 2 7B training task, the system achieved 6.584 minutes to train, while on the DeepSeek 671B ultra-large model test — a workload that pushes memory and interconnect bandwidth to their limits — it logged 84.85. The system also posted 74.75 on Llama 3.1 8B and 2.342 on the DLRM-DCN recommendation benchmark, both of which test different axes of the AI infrastructure stack. The DLRM-DCN score is particularly useful for enterprises running recommendation engines at scale, which is one of the most common production AI workloads in retail and ad-tech.

The submission reflects ongoing participation by ASUS in MLCommons, the industry body that maintains MLPerf as the de facto standard for AI performance benchmarking. By contributing results and continuously optimising its server platforms, ASUS is positioning itself as a long-term player in the enterprise AI infrastructure market, where benchmark transparency is increasingly a procurement requirement. The HGX B300 is NVIDIA’s top-tier Blackwell Ultra offering, and ASUS is one of a small set of OEMs shipping systems built on it at scale.

Availability and Enterprise Positioning

The ASUS XA NB3I-E12 is shipping worldwide, with availability through ASUS’s enterprise and channel partner network. Pricing is not publicly listed because enterprise AI servers are typically sold with configuration customisation, support contracts, and deployment services bundled in. Customers looking to procure the system are directed to contact their local ASUS representative for configuration and quotation. This is the standard sales model for flagship AI servers in this class — buyers are not picking boxes off a shelf, they are working with OEMs on rack-level integration, networking, and validation.

For Malaysian buyers, the relevant route is through ASUS Malaysia’s commercial and enterprise sales team, which handles AI infrastructure deployments for data centre operators, government research clusters, and large enterprise training facilities. The XA NB3I-E12 is positioned against systems from Dell PowerEdge, HPE Cray, Lenovo ThinkSystem, and Supermicro in the HGX B300 segment, and the MLPerf #1 rankings give ASUS a clear marketing hook for the workloads that matter most to enterprise buyers.

Our Take

The MLPerf #1 sweep is a meaningful data point, but the real story is the breadth of the wins. Llama 3.1-8B Server mode covers multi-user serving, Offline mode covers batch throughput, and GPT-OSS-120B Interactive covers reasoning latency on the largest models. Winning all three in the same submission is unusual — most vendors top one or two categories while losing ground on the third. For enterprises running mixed AI workloads, the XA NB3I-E12 now has a credible case as a single platform that can handle training, fine-tuning, and inference without compromising on any axis.

The HGX B300 platform itself is a significant generational leap over HGX H100, and ASUS is shipping early — which is what you want from a vendor if you are running time-sensitive AI deployments. The eight-Blackwell-Ultra configuration is at the top of the SKU stack, so buyers evaluating the XA NB3I-E12 should expect pricing to reflect that positioning, with lower-tier configurations available within the same platform family. Our recommendation: for organisations that need MLPerf-proven performance and a vendor actively contributing benchmark results, the XA NB3I-E12 belongs on the shortlist. For smaller workloads that do not need eight GPUs and full HGX B300 bandwidth, look at lower-tier ASUS AI server SKUs that share the firmware and management tooling but come in smaller configurations. This is a data centre story rather than a consumer one, and it sits alongside the Kingston AI solutions coverage from COMPUTEX 2025 as part of the broader enterprise AI infrastructure conversation we have been tracking.

• ASUS AI Servers https://www.asus.com/commercial-servers-workstations/
• MLCommons MLPerf Training v6.0 Results https://mlcommons.org/benchmarks/mlperf-training/
• NVIDIA HGX B300 Platform https://www.nvidia.com/en-us/data-center/hgx/