移至主內容

Networking Hardware, Devices and Equipment for Data Centers

Supermicro Delivers Accelerated Ethernet Networking Optimized for AI With NVIDIA Spectrum™-X

AI is driving the need for more advanced compute infrastructure, including networking. A lack of networking bandwidth often creates a bottleneck in the information processing that drives AI applications. Accelerated computing infrastructure is only as fast as its weakest link, and all too often, that's networking. Fortunately, by working with NVIDIA, Supermicro has qualified the new NVIDIA Spectrum-X networking platform in several of its systems to allow for up to 1.6 times better networking performance for AI workloads, along with increased energy efficiency and reduced total cost of ownership. With the combination of the NVIDIA Spectrum-4 Ethernet switch and the NVIDIA BlueField®-3 SuperNIC, customers can now take advantage of the highest-performing Ethernet AI cloud processing infrastructure.

NVIDIA Spectrum-4 and NVIDIA BlueField-3

Systems now being qualified with the NVIDIA Spectrum-X networking solution include Supermicro's very popular GPU SuperServers that come in 4U, 5U, and 8U heights and support NVIDIA GPUs in the following systems:

Trio of Supermicro GPU systems: 4U 8GPU, 8U 8GPU, and 5U 10GPU

These Supermicro systems, powered by NVIDIA Spectrum-X, reduce the training and inference runtimes of massive transformer-based generative AI models, essential for delivering faster results and insights from AI.

In addition to tremendously increasing the network's performance by 1.6 times for AI, Supermicro systems with NVIDIA Spectrum-X networking benefit from isolating workload performance in multi-tenant environments. Allowing workloads to move data separately and independently, free from "noisy-neighbor" interference that could impede performance, is critical for high-performance applications. Additionally, isolating network workloads provides an added security benefit of preventing nefarious actors in one workload from interfering with a separate workload. NVIDIA Spectrum-X accomplishes this enhanced workload isolation through secure virtual private clouds (VPCs) assigned to each tenant.

Energy efficiency has always been a critical concern when operating high-performance, high-power AI systems. By delivering increased networking performance within the same networking power envelope, NVIDIA Spectrum-X can contribute to a more power-efficient AI environment. This better performance per watt leads to lower overall power consumption and potentially lower total cost of ownership. For users that commonly rely on power capping to keep control of energy costs, Supermicro systems with Spectrum-X help provide reassurance that energy consumption caps are not surpassed.  

Consumers who have standardized Ethernet networking can be assured that NVIDIA Spectrum-X is fully standards-based and completely interoperable with Ethernet-based stacks. Additionally, Supermicro servers with NVIDIA Spectrum-X help customers overcome the most significant challenge of deploying AI—scalability. With the ability to support 256 200G ports in a single switch or 16K ports in a two-tier leaf/spine topology, these are ideal systems for organizations with evolving accelerated compute infrastructure needs.

As Supermicro strives to meet the needs of AI customers who need not only high-performance networking but also cloud multi-tenancy, Supermicro systems with NVIDIA Spectrum-X leverage remote direct memory access (RDMA) over Converged Ethernet (RoCE) extensions to support both performance and multi-tenancy. The performance over traditional Ethernet is astonishing, as illustrated by the chart below.  

NVIDIA's breakthrough networking technology, including in numerous Supermicro servers, achieves breakthrough network performance for AI, multi-tenant performance isolation, and increased energy efficiency. All this is done while building on Ethernet networking standards by coupling the NVIDIA Spectrum-4 Ethernet switch with the NVIDIA BlueField-3 SuperNIC. This combination has the net effect of helping reduce the runtimes of massive generative AI models. For customers who need faster time to critical insights, multi-tenancy, and future scalability, these Supermicro systems with NVIDIA Spectrum-X are ideally suited to meet their needs.