NVIDIA has introduced the NVIDIA MGX server specification, which is designed to meet the diverse accelerated computation needs of data centers around the globe. This modular reference architecture would enable system manufacturers to rapidly and cost-effectively construct over 100 server variants to support a vast array of applications, including artificial intelligence (AI), high-performance computing (HPC), and Omniverse workloads.
ASRock Rack, ASUS, GIGABYTE, Pegatron, QCT, and Supermicro, among others, have announced their adoption of the MGX server specification. By utilizing MGX, these manufacturers can reduce development costs by as much as 75% and reduce development time to just six months.
“Enterprises are looking for more options for accelerated computing when designing data centers to meet their unique business and application requirements,” said Kaustubh Sanghani, Vice President of GPU Products at NVIDIA. “We developed MGX to assist organizations in implementing enterprise AI while saving them significant time and money.”
The MGX server specification permits manufacturers to begin with a system architecture optimized for accelerated computing and then select the appropriate GPU, DPU, and CPU for their particular needs. This adaptability enables design variations to accommodate unique applications, such as high-performance computing, data science, large language models, periphery computing, graphics and video processing, enterprise artificial intelligence, and design and simulation. In addition to facilitating the management of multiple duties, such as AI training and 5G, on a single machine, the modular approach would ensure seamless hardware enhancements for future generations. MGX is readily compatible with cloud and enterprise data centers.
Collaboration with QCT, Supermicro, and SoftBank
QCT and Supermicro will be among the first companies to introduce MGX-based designs to the market in August 2023. The ARS-221GL-NR system from Supermicro will feature the NVIDIA Grace CPU Superchip, while the S74G-2U system from QCT will employ the NVIDIA GH200 Grace Hopper Superchip.
SoftBank Corporation intends to leverage MGX in its numerous hyperscale data centers across Japan, in addition to server manufacturers. The company intends to dynamically allocate GPU resources between generative AI and 5G applications in order to address the challenges posed by constructing a cost-effective infrastructure.
Meeting Specific Demands and Maximizing Adaptability
Data centers must balance the increasing demand for computing power with the need to reduce their ecological footprint and operating expenses. Since their inception, NVIDIA’s accelerated computation servers have provided superior performance and energy efficiency, according to NVIDIA. With the introduction of the MGX server specification, system manufacturers are now able to tailor each server to the budget, power delivery, thermal design, and mechanical requirements of individual clients.
MGX is compatible with current and future generations of NVIDIA hardware, including various chassis options (1U, 2U, and 4U with air or liquid cooling), a full range of NVIDIA GPUs, NVIDIA Grace and GH200 Grace Hopper Superchips, x86 CPUs, as well as NVIDIA BlueField-3 DPUs and ConnectX-7 network adapters.
MGX is distinct from NVIDIA HGX
MGX offers flexible, cross-generational compatibility with NVIDIA products, whereas NVIDIA HGX is designed for AI and HPC systems with scalable multi-GPU configurations connected via NVLink. This compatibility would enable systems integrators to utilize existing designs and implement next-generation products without costly redesigns.
Support for all Software Features
NVIDIA’s extensive software infrastructure supports the MGX server specification, empowering developers and enterprises to construct and accelerate AI, HPC, and other applications. This includes NVIDIA AI Enterprise, the software element of the NVIDIA AI platform, which offers over one hundred frameworks, pretrained models, and development tools for accelerated AI and data science.
