NVIDIA’s Anthony Larijani, Neil Dey, Ivan Goldwasser, Itay Ozery, and Michael Mangiafico have unveiled a 6U (800 mm) chassis configuration built upon the MGX modular reference architecture, designed for next-generation accelerated compute and networking platforms. This design supports both current x86 platforms and the forthcoming NVIDIA Vera CPU, enabling standardized building blocks for flexible, energy-efficient AI data centers. The MGX 6U chassis features a liquid-cooled variant of the NVIDIA RTX PRO 6000 Blackwell Server Edition GPU, alongside NVIDIA BlueField-3 DPUs and ConnectX-8 SuperNICs with PCIe Gen 6, aiming to address growing demands for power, thermal capacity, and facility space in rapidly expanding AI deployments.
MGX 6U Modular Architecture for Flexible AI Data Centers
The NVIDIA MGX 6U modular architecture provides a flexible foundation for building future-ready AI data centers. This design supports multiple computing generations, including both x86 platforms and the upcoming NVIDIA Vera CPU, simplifying integration for system partners. The 6U chassis (800mm) prioritizes serviceability with accessible pathways for key components like network cards and power supplies, reducing operational overhead when managing large-scale infrastructure.
The MGX 6U design is central to the new liquid-cooled NVIDIA RTX PRO Server, featuring eight RTX PRO 6000 Blackwell Server Edition GPUs. This configuration utilizes NVIDIA ConnectX-8 SuperNICs with PCIe Gen 6, capable of delivering up to 400 Gb/s of bandwidth per GPU. Integrated PCIe Gen 6 switches within the ConnectX-8 effectively double per-GPU network bandwidth and streamline design, removing I/O bottlenecks for improved performance in multi-GPU workloads.
Enhanced security and infrastructure acceleration are achieved through the inclusion of NVIDIA BlueField DPUs within the MGX 6U architecture. These DPUs offload functions like encryption and threat detection, securing AI pipelines and freeing GPU/CPU resources for core AI tasks. Validated BlueField-accelerated applications further enhance efficiency and cybersecurity coverage, enabling consistent policy enforcement across the AI factory infrastructure.
Enhanced AI Performance with NVIDIA RTX PRO 6000 Blackwell
The NVIDIA MGX 6U design supports up to eight liquid-cooled RTX PRO 6000 Blackwell Server Edition GPUs, delivering efficient performance for next-generation AI factories. This new server configuration also features NVIDIA BlueField-3 DPUs and ConnectX-8 SuperNICs with PCIe Gen 6, maximizing performance for enterprise AI workloads. Validated compatibility with over 50 ISVs ensures broad software support for engineering, scientific computing, and professional visualization applications.
Enhanced AI networking is achieved through ConnectX-8 SuperNICs supporting up to 400 Gb/s per RTX PRO 6000 Blackwell GPU, effectively doubling per-GPU network bandwidth. Integrated PCIe Gen 6 switches streamline the design and reduce server complexity, removing I/O bottlenecks and speeding data movement. This results in up to 2x higher NCCL all-to-all performance, supporting more scalable multi-GPU, multi-node workloads across AI factories.
NVIDIA BlueField DPUs within the MGX 6U design provide zero-trust security and accelerate infrastructure functions. These DPUs offload tasks like encryption and threat detection, freeing GPU/CPU resources for AI workloads. By isolating control and management planes, BlueField enhances security and accelerates networking, storage, and virtualization services, enabling consistent performance and policy enforcement across AI factory deployments.
High-Performance AI Networking with NVIDIA ConnectX
High-performance AI networking is a key feature of the MGX 6U design, utilizing NVIDIA ConnectX-8 SuperNICs. These SuperNICs, with integrated PCIe Gen 6 switches, support up to 400 Gb/s of network bandwidth per RTX PRO 6000 Blackwell GPU. This effectively doubles per-GPU network bandwidth compared to systems with dedicated switches, removing I/O bottlenecks and accelerating data movement for multi-GPU workloads. The MGX 6U also supports future upgrades to ConnectX-9, providing Ethernet and InfiniBand connectivity options.
The MGX 6U reference design streamlines system architecture by integrating the latest generation ConnectX-8 SuperNICs directly. This simplifies design and reduces server complexity, while maximizing network performance for demanding AI applications. The integrated PCIe Gen 6 switches within the ConnectX-8 SuperNICs contribute to up to 2x higher NCCL all-to-all performance, making it ideal for scalable multi-GPU, multi-node AI factory deployments.
NVIDIA BlueField DPUs work alongside ConnectX-8 to enhance AI infrastructure. BlueField delivers zero-trust security and accelerates infrastructure functions like encryption and threat detection, offloading these tasks from the GPU/CPU. This allows the RTX PRO Server to focus on AI workloads while maintaining robust security and consistent policy enforcement across the AI factory, ensuring scalable and secure deployments.
Built for AI workloads, ConnectX-8 with integrated PCIe Gen 6 switches supports up to 400 Gb/s of network bandwidth per RTX PRO 6000 Blackwell GPU (based on a 2:1 GPU-to-NIC ratio).
AI Runtime Security and Acceleration with NVIDIA BlueField
The NVIDIA MGX 6U design incorporates NVIDIA BlueField DPUs to enhance AI runtime security and accelerate infrastructure functions. These DPUs deliver zero-trust security by offloading tasks like line-rate encryption and micro-segmentation. This protects AI pipelines from threats while preserving GPU/CPU resources for core AI workloads. BlueField also accelerates networking, storage, and virtualization services, ensuring consistent performance and policy enforcement across AI factory deployments.
NVIDIA BlueField DPUs within the MGX 6U architecture contribute to infrastructure acceleration by offloading critical functions. This allows organizations to isolate control and management planes in hardware, improving security and efficiency. Validated BlueField-accelerated applications from software providers further extend these capabilities, boosting both infrastructure efficiency and cybersecurity coverage for scalable AI deployments.
The MGX 6U design, featuring BlueField, is intended to build future-ready AI factories. By incorporating these DPUs, the architecture ensures consistent performance and policy enforcement, supporting scalable multi-GPU, multi-node workloads. This approach allows AI factories to evolve alongside silicon innovations, reducing engineering costs and shortening time-to-market for system partners.
Sustainable Computing with Liquid-Cooling Technology
The NVIDIA MGX 6U design supports sustainable computing through liquid-cooling technology, specifically with the new liquid-cooled NVIDIA RTX PRO 6000 Blackwell Server Edition GPU. This configuration features eight of these GPUs within a single chassis, improving thermal efficiency and allowing for compact, single-slot form factors. The liquid cooling is foundational for the next wave of accelerated computing platforms, enabling improved performance for demanding enterprise AI workloads and AI factories.
High-performance AI networking is achieved within the MGX 6U design through NVIDIA ConnectX-8 SuperNICs, supporting up to 400 Gb/s per GPU with integrated PCIe Gen 6 switches. This effectively doubles per-GPU network bandwidth compared to systems with dedicated switches, removing I/O bottlenecks and accelerating data movement. The design supports both current and future networking technologies, including the forthcoming ConnectX-9, offering Ethernet and InfiniBand connectivity options for diverse data center requirements.
NVIDIA BlueField DPUs integrated into the MGX 6U design enhance security and infrastructure acceleration. These units offload functions like encryption and threat detection, enforcing zero-trust security while freeing GPU/CPU resources for AI workloads. By isolating control and management planes, BlueField protects AI pipelines and accelerates networking, storage, and virtualization, ensuring scalable and secure deployments for AI factories and consistent policy enforcement.
With a compact, single-slot liquid-cooled form factor, RTX PRO 6000 Blackwell delivers breakthrough performance for powering AI factories and accelerating demanding enterprise AI workloads with improved thermal efficiency.
