Eight NVIDIA H200 NVL GPUs in a 4U Supermicro-built PCIe server, sold and supported by Cisco and managed through Cisco Intersight, for enterprise AI inference without rack-scale complexity.

The AI Server SYS-422GL-FNR2 is a 4RU air cooled MGX PCIe GPU server for large model inference and fine tuning. Configured with eight NVIDIA H200 NVL Tensor Core GPUs, it provides 1,128 GB of HBM3e memory and 38.4 TB/s of aggregate GPU memory bandwidth in a single node.

Memory capacity is the reason to choose this platform. Each H200 NVL carries 141 GB of HBM3e. With optional NVLink bridging, four GPUs form a single 564 GB memory domain, large enough to hold and fine tune model classes that would otherwise be sharded across multiple servers. Sharding costs interconnect bandwidth, scheduling complexity, and engineering time. Keeping a model resident in one domain avoids all of it. 

H200 NVL is a PCIe dual slot, air cooled accelerator. Organizations can use existing air-cooling for simplified deployment without the requirements of liquid cooling infrastructure. The Supermicro SYS-422GL-FNR2 Networking is handled by an NVIDIA ConnectX-8 PCIe switch board, which combines a 400/800 Gb/s SuperNIC with an integrated PCIe Gen6 switch. Collapsing those two functions into a single component removes a hop between GPU and fabric and frees up to eight PCIe 6.0 x16 double-width slots for accelerators.

Two Intel Xeon processors, up to 3 TB of DDR5 memory, and a dual root PCIe 5.0 switch topology keep the accelerators supplied. NVIDIA ConnectX-8 SuperNICs handle east-west traffic and NVIDIA BlueField-3 handles storage and north-south connectivity. Cisco Intersight provides lifecycle management across the fleet, and CX Services provides support.

Table 1 summarizes the primary features and benefits of the platform.

Table 1. Features and benefits 

Dual-root PCIe architecture with NVIDIA ConnectX-8 

Deployment topologies

The dual root PCIe 5.0 switch supports two configurations, selected at order time by whether the NVLink bridge kit is included.

Without bridging, the system presents eight independent GPUs. Each is an addressable inference endpoint with 141 GB of HBM3e. This configuration maximizes tenant density and replica count, and suits high concurrency serving of models that fit within a single GPU.

With 4-way bridging, the system presents two 564 GB memory domains. This configuration suits fine tuning and inference of models that exceed single GPU memory, and it keeps those models inside one server rather than distributing them across a cluster.

On network sizing, the platform carries four ConnectX-8 SuperNICs across eight GPUs. This ratio is designed for inference serving and single node fine tuning. Customers with multi-node distributed training requirements should evaluate the Cisco Supermicro HGX B300 platforms, which provide a one to one GPU to SuperNIC ratio for scale out fabrics. 

Multi-tenancy and workload isolation 

Multi-Instance GPU divides each H200 NVL into as many as seven instances, each with dedicated memory, cache, and streaming multiprocessors. A fully populated server therefore supports up to 56 concurrent isolated workloads. Because the partitioning is enforced in hardware, a fault or a resource spike in one instance does not affect the others, which makes quality of service predictable in shared environments.  

NVIDIA Confidential Computing is supported on the Hopper architecture. Data and model weights remain encrypted while in use, protected from the hypervisor, the host operating system, and other tenants on the same physical server. For regulated industries, sovereign deployments, and service providers who must demonstrate isolation rather than assert it, this is often the requirement that determines platform selection. 

Software and management

NVIDIA AI Enterprise is included with each H200 NVL GPU. It provides supported inference runtimes, NVIDIA NIM microservices for deploying models as API endpoints, and validated builds of common training and inference frameworks.  

Cisco Intersight manages the platform through its lifecycle, covering server profiles, firmware, configuration policy, hardware telemetry, and health monitoring. Servers can be managed individually or as a pool, and Intersight applies consistent policy across select Supermicro server deployments.

Most PCIe GPU servers bottleneck at the host bridge. This system uses a dual-root design: each Xeon CPU owns a PCIe 5.0 x16 path into the NVIDIA ConnectX-8 switch board, which fans out to the GPU slots and integrated QSFP networking at up to 400 Gb/s. GPUto-GPU traffic stays on the switch board instead of crossing the CPU interconnect, and optional NVLink bridges pair adjacent GPUs for workloads that need more than PCIe bandwidth. 

The practical result: eight GPUs behave like a coherent inference pool, not eight isolated cards. For scale-out, four ConnectX-8 SuperNICs carry east-west traffic and up to two NVIDIA BlueField-3 B3220 Crypto controllers offload storage and north-south networking from the host CPUs.

Cisco offers this Supermicro platform in a defined configuration, listed in Table 2.

Table 2. Cisco selected configuration, SYS-422GL-FNR2 (MGX H200)

This is a hardware product; no platform license is required. Cisco Intersight management requires an Intersight license, listed in Table 3.

Table 3. Licensing

Information about Cisco’s Environmental, Social and Governance (ESG) initiatives and performance is provided in Cisco’s CSR and sustainability reporting.

Table 4. Cisco Environmental Sustainability Information

Table 5 reflects the Supermicro data sheet for the SYS-422GL-FNR2. Product specifications may change without notice.

Table 5. Product specifications, Supermicro GPU SuperServer
SYS-422GL-FNR2

Table 6 lists what must be in place before installation.

Table 6. System requirements 

The base platform is a Supermicro SKU offered through Cisco. Cisco ordering PIDs for the configured system are in progress. To order, contact your Cisco account representative or visit the Cisco Ordering Home Page.

Table 7. Ordering information

Warranty information

Warranty terms for this Cisco-offered Supermicro platform are standard Supermicro warranty terms. Support is delivered through Cisco CX Services under the applicable service contract, contact your Cisco sales specialist for more details.

Cisco and Partner Services

Cisco CX and Supermicro Services cover this platform from planning through deployment and ongoing operations, with dedicated support contracts For more information, visit https://www.cisco.com/go/services.

Cisco Capital

Flexible payment solutions to help you achieve your objectives

Cisco Capital makes it easier to get the right technology to achieve your objectives, enable business transformation and help you stay competitive. We can help you reduce the total cost of ownership, conserve capital, and accelerate growth. In more than 100 countries, our flexible payment solutions can help you acquire hardware, software, services and complementary third-party equipment in easy, predictable payments. Learn more.

Learn more

For more on Cisco AI infrastructure, visit https://www.cisco.com/go/ai. Full platform specifications are available on the Supermicro product page for the SYS-422GL-FNR2.

Learn more about Cisco and Supermicro partnership:
https://www.cisco.com/site/us/en/solutions/global-partners/supermicro/index.html.