NVIDIA HGX is at the heart of many of today's most powerful AI and high-performance computing servers!

Unlike NVIDIA DGX, which is delivered as a complete NVIDIA system, NVIDIA HGX is a computing platform that server manufacturers integrate into their own AI server architectures.

Dell Technologies, HPE, Lenovo, Supermicro, Cisco, GIGABYTE and other manufacturers build AI servers around NVIDIA HGX platforms, including HGX H100, HGX H200, HGX B200 and HGX B300.

These high-density AI systems place exceptional demands on rack power distribution. Power requirements, the number of power supplies, connector types, redundancy and power architecture can differ significantly between server models and configurations.

Schleifenbauer designs and manufactures build-to-order Rack PDUs for NVIDIA HGX-based servers, regardless of the server manufacturer. The PDU can be configured around the actual server model, power supplies, A/B redundancy, rack layout and electrical infrastructure of the data center.

What is NVIDIA HGX?

NVIDIA HGX is an accelerated computing platform used by server manufacturers to build high-performance AI systems. HGX combines multiple NVIDIA SXM GPUs with technologies including NVIDIA NVLink and NVSwitch to provide high-speed GPU-to-GPU communication.

As a result, the same NVIDIA HGX platform can be found inside AI servers from many different manufacturers.

Examples include:

Manufacturer	Examples of HGX systems	NVIDIA platform
Dell Technologies	PowerEdge XE9680	HGX H100 / H200
Dell Technologies	PowerEdge XE9680L, XE9780	HGX B200
HPE	Cray XD670	HGX H100 / H200
HPE	Compute XD690	HGX B300
Lenovo	ThinkSystem SR680a V3	HGX H100 / H200 / B200
Lenovo	ThinkSystem SR780a V3	HGX H100 / H200 / B200
Lenovo	ThinkSystem SR680a V4	HGX B300
Supermicro	Various GPU SuperServer systems	HGX H100 / H200 / B200 / B300
Cisco	UCS C885A M8	HGX H100 / H200
Cisco	UCS C880A M8	HGX B300
GIGABYTE	G593, G893, G894 and G4L4 series	HGX H100 / H200 / B200 / B300

NVIDIA-Certified HGX systems are also available from manufacturers including ASUS, Eviden, Inventec, Pegatron, QCT, Wistron and others.

Available configurations continue to change as new generations of GPUs and AI servers are introduced. For this reason, Schleifenbauer always starts with the actual server model and its electrical specifications when configuring a Rack PDU.

Why NVIDIA HGX AI servers require a different approach to rack power

A modern HGX server can concentrate an extremely high power demand into only a few rack units. These systems can also contain multiple high-power redundant power supplies.

This changes the requirements placed on the Rack PDU.

In traditional IT infrastructure, a PDU may primarily be selected according to the total rack load and number of required outlets. With high-density AI infrastructure, many more factors need to be considered together:

• maximum and expected power consumption per server;
• number of power supplies;
• power connector type;
• A/B power feeds and redundancy;
• phase load distribution;
• number of HGX systems per rack;
• available three-phase power;
• circuit protection;
• cable management and available rack space;
• input, branch and outlet-level power monitoring;
• future expansion of the AI rack.

This is why there is no single universal PDU for every NVIDIA HGX system.

The correct PDU depends on the complete server and rack configuration.

Request a PDU quote directly

PDU for Dell PowerEdge NVIDIA HGX servers

Dell Technologies integrates NVIDIA HGX into several systems within its PowerEdge portfolio.

NVIDIA-Certified systems include, for example, the Dell PowerEdge XE9680 with NVIDIA HGX H100 or H200, as well as the Dell PowerEdge XE9680L and XE9780 with NVIDIA HGX B200.

For these systems, Schleifenbauer can configure the Rack PDU according to factors including:

• the exact Dell PowerEdge model;
• number and capacity of the power supplies;
• required high-power outlets;
• A/B power feeds;
• three-phase power distribution;
• required redundancy;
• input, branch or outlet-level monitoring.

This means the power distribution can be designed around the Dell AI server rather than adapting the server infrastructure to a standard PDU.

PDU for HPE NVIDIA HGX systems

HPE also offers high-density AI and HPC systems based on NVIDIA HGX.

Examples include systems from the HPE Cray and HPE Compute portfolios using NVIDIA HGX H100, H200 and B300 platforms.

The HPE Compute XD690 with NVIDIA HGX B300, for example, demonstrates the power density that can be achieved with the latest generation of AI infrastructure.

For HPE HGX systems, Schleifenbauer can design a Rack PDU configuration around the electrical infrastructure of the data center, the server power supplies, rack layout and required A/B redundancy.

PDU for Lenovo ThinkSystem NVIDIA HGX servers

Lenovo uses NVIDIA HGX across several ThinkSystem GPU servers.

Examples include:

• Lenovo ThinkSystem SR680a V3 with HGX H100, H200 or B200;
• Lenovo ThinkSystem SR780a V3 with HGX H100, H200 or B200;
• Lenovo ThinkSystem SR680a V4 with HGX B300.

These systems illustrate why total power consumption is only part of the PDU specification. The number of physical power supplies, their capacity and their distribution across redundant feeds must also be considered.

A Schleifenbauer Rack PDU can be configured specifically around the ThinkSystem server architecture and the available rack power infrastructure.

PDU for Supermicro NVIDIA HGX GPU servers

Supermicro offers an extensive portfolio of high-density GPU servers supporting multiple generations of NVIDIA HGX.

NVIDIA-Certified Supermicro systems are available for HGX H100, HGX H200, HGX B200 and HGX B300.

For data centers deploying Supermicro AI infrastructure, Schleifenbauer can configure the Rack PDU around the exact Supermicro server model, number of GPU nodes per rack and available electrical infrastructure.

This applies to individual AI servers as well as racks containing multiple GPU systems with very high combined power requirements.

PDU for Cisco, GIGABYTE and other NVIDIA HGX servers

The Schleifenbauer PDU architecture is not tied to one server manufacturer.

Rack PDUs can be configured for HGX-based systems from manufacturers including:

Dell Technologies, HPE, Lenovo, Supermicro, Cisco, GIGABYTE, ASUS, Eviden, Inventec, Pegatron, QCT and Wistron.

If your server manufacturer or model is not listed here, this does not mean Schleifenbauer cannot provide a suitable PDU.

Using the electrical specifications of the server and rack infrastructure, our engineers can determine and configure the required power distribution solution.

From NVIDIA HGX H100 and H200 to B200 and B300

AI hardware is developing rapidly.

NVIDIA HGX H100 and H200 systems are now joined by Blackwell platforms such as HGX B200 and Blackwell Ultra HGX B300. At the same time, AI infrastructure continues to increase rack power density.

This means power infrastructure needs to account not only for today's servers, but also for factors such as:

• increasing rack power requirements;
• more high-power connections;
• higher loads per circuit;
• redundant power feeds;
• accurate energy monitoring;
• scalability towards future AI hardware.

A build-to-order Rack PDU allows power distribution to be designed around the infrastructure rather than selecting from a limited range of standard configurations.

32A and 63A three-phase PDUs for high-density AI racks

For high-density AI and HPC environments, Schleifenbauer can configure 32A and 63A three-phase Rack PDUs, as well as other power configurations depending on the application.

Different input connections, outlet configurations, circuit protection and monitoring options are available.

The appropriate configuration can be determined according to:

Server model → number of power supplies → maximum power → redundancy → available rack power → required monitoring.

This allows the PDU to be configured for a single HGX server, multiple servers within one rack or a larger AI cluster architecture.

A/B power redundancy for NVIDIA HGX systems

Availability is critical for AI training, inference and HPC workloads.

Where a server has redundant power supplies, these can be distributed across separate A and B power feeds. The power architecture can therefore be designed so that the loss of one feed does not automatically remove power from the complete server.

Schleifenbauer can design PDU configurations taking into account:

• A and B power feeds;
• PSU distribution;
• phase balancing;
• circuit loading;
• redundancy;
• available capacity per PDU.

The configuration is always based on the actual server and data center power architecture.

Power monitoring for AI and GPU infrastructure

As the power density of NVIDIA HGX systems increases, visibility into energy consumption becomes increasingly important.

Schleifenbauer PDU 5.0 is available as:

• Basic PDU;
• Metered PDU;
• Monitored PDU;
• Switched PDU;
• Managed PDU.

Depending on the configuration, electrical parameters can be measured at input, branch and outlet level.

This provides visibility into how much power an AI server or power circuit is actually consuming and how much capacity remains available within the rack.

NVIDIA HGX power monitoring with EnerTree DCEM

Intelligent Schleifenbauer PDUs can be combined with EnerTree DCEM, Schleifenbauer's Data Center Energy Management platform.

EnerTree provides central visibility into rack-level energy data and supports the monitoring and analysis of data center power infrastructure.

This is particularly relevant in AI and HPC environments, where increasing rack power densities make accurate monitoring of load, capacity and energy consumption increasingly important.

EnerTree supports open integration through protocols and interfaces including REST API, SNMP, Modbus and Redfish.

The software is included within the Schleifenbauer PDU architecture without recurring software license fees and can be deployed on-premises, giving operators control over their own energy data.

Build-to-order PDU instead of adapting the server to the PDU

An NVIDIA HGX server represents a significant infrastructure investment. The rack power distribution should therefore match the server architecture rather than forcing the infrastructure to work around a standard PDU.

Schleifenbauer designs and manufactures Rack PDUs build-to-order in the Netherlands.

Depending on the application, configurations can include:

• single-phase or three-phase power;
• 16A, 32A or 63A;
• required number of outlets;
• C13, C19 and IEC Lock CX outlets;
• cable length;
• input connector;
• power cable position;
• circuit protection;
• vertical 0U or horizontal mounting;
• input, branch and outlet-level metering;
• remote outlet switching;
• communication and monitoring.

The PDU can therefore be configured around both the AI server and the complete rack power architecture.

Click here to request a quote

Which PDU do I need for my NVIDIA HGX server?

That depends on the exact system.

A Dell PowerEdge server with NVIDIA HGX B200 can have different power supplies, connectors and electrical requirements from a Lenovo ThinkSystem, HPE Compute or Supermicro server based on the same NVIDIA HGX platform.

To determine the appropriate Rack PDU, the following information is particularly useful:

1. server manufacturer;
2. exact server model;
3. NVIDIA HGX platform or GPU configuration;
4. number of servers per rack;
5. number and type of power supplies per server;
6. required A/B redundancy;
7. available rack power;
8. required power and energy monitoring.

Based on this information, Schleifenbauer can configure a Rack PDU that matches the actual AI infrastructure.

Frequently asked questions about PDUs for NVIDIA HGX

What is the difference between NVIDIA DGX and NVIDIA HGX?

NVIDIA DGX is a complete AI computing system from NVIDIA. NVIDIA HGX is an accelerated computing platform that server manufacturers integrate into their own AI servers. This means HGX-based systems are available from manufacturers including Dell Technologies, HPE, Lenovo, Supermicro, Cisco and GIGABYTE.

Which server manufacturers use NVIDIA HGX?

NVIDIA HGX is used in servers from multiple manufacturers. NVIDIA-Certified HGX systems include models from Dell Technologies, HPE, Lenovo, Supermicro, Cisco, GIGABYTE, ASUS, Eviden, Inventec, Pegatron, QCT, Wistron and others. Availability varies between HGX generations and individual server platforms.

Which PDU do I need for NVIDIA HGX H100?

The correct PDU for an NVIDIA HGX H100 system depends on the server manufacturer and model, number and capacity of the power supplies, connector type, available rack power and required redundancy. Schleifenbauer can configure the PDU according to the electrical requirements of the specific HGX H100 server.

Which PDU do I need for NVIDIA HGX H200?

There is no single PDU specification for every HGX H200 server. Dell, HPE, Lenovo, Supermicro and other manufacturers can use different power architectures around the HGX platform. The Rack PDU should therefore be selected according to the actual server configuration.

Which PDU do I need for NVIDIA HGX B200?

There is no universal PDU for HGX B200. The appropriate PDU depends on the server manufacturer, server model, number and type of power supplies, available supply, A/B redundancy and number of servers within the rack. Schleifenbauer can configure the PDU specifically for these requirements.

Which PDU do I need for NVIDIA HGX B300?

HGX B300 is available in systems from multiple manufacturers, including Dell, HPE, Lenovo, Supermicro, Cisco and GIGABYTE. The correct PDU configuration depends on the electrical architecture of the specific server and rack rather than the HGX B300 platform alone.

Can Schleifenbauer provide a PDU for Dell PowerEdge HGX servers?

Yes. Schleifenbauer can configure Rack PDUs according to the electrical requirements of Dell PowerEdge AI servers with NVIDIA HGX, including systems from the PowerEdge XE portfolio. The final PDU configuration is based on the exact server model and rack power architecture.

Can Schleifenbauer provide a PDU for HPE NVIDIA HGX systems?

Yes. Rack PDUs can be configured for HPE AI and HPC systems with NVIDIA HGX, taking into account the power supplies, connections, rack power, A/B redundancy and monitoring requirements of the specific server.

Can Schleifenbauer provide a PDU for Lenovo ThinkSystem HGX servers?

Yes. This includes Lenovo ThinkSystem servers using NVIDIA HGX H100, H200, B200 and B300. The PDU is configured according to the specific ThinkSystem model and electrical infrastructure.

Can Schleifenbauer provide a PDU for Supermicro HGX servers?

Yes. Schleifenbauer can configure Rack PDUs for Supermicro GPU servers based on NVIDIA HGX H100, H200, B200 and B300, depending on the specifications of the individual server model.

Are 63A PDUs suitable for NVIDIA HGX and AI racks?

63A three-phase Rack PDUs can be suitable for high-density AI and HPC racks. However, the appropriate capacity should always be determined from the actual server load, electrical infrastructure, redundancy requirements and number of systems within the rack.

Can I monitor the power consumption of NVIDIA HGX servers?

Yes. With an appropriate Schleifenbauer Monitored or Managed PDU, electrical parameters can be measured at different levels. EnerTree DCEM can then be used to centrally monitor and analyze this energy data.

Can the PDU be configured for A/B redundant power feeds?

Yes. Where the server architecture uses redundant power supplies, Schleifenbauer can configure the PDU solution around separate A and B feeds, taking into account PSU distribution, phase balancing, circuit loading and the required redundancy.

Need a PDU for an NVIDIA HGX AI server?

Whether you are deploying a Dell PowerEdge, HPE, Lenovo ThinkSystem, Supermicro, Cisco, GIGABYTE or another NVIDIA HGX-based system, the rack power distribution needs to match the actual server and infrastructure.

Send us the server manufacturer and model, number of systems per rack and available power supply. Our engineers can use this information to configure a suitable Rack PDU solution.

HGX H100, H200, B200 or B300. One AI server or a complete GPU cluster. The PDU is built around your infrastructure.

The post PDUs for NVIDIA HGX AI Servers appeared first on Schleifenbauer – Bespoke PDUs.

Source: https://www.schleifenbauer.eu/en/nvidia-hgx-pdu/