AMD Helios: Rack Design, Power and Cooling Requirements

AMD Helios is a rack-scale reference design whose deployed configuration comes from an OEM or ODM. For facility planning, AMD and Schneider Electric’s August 2026 reference design supports racks up to 246 kW. Treat that as a reference-design envelope: the power, cooling, dimensions and service requirements for a purchased Helios system must come from its specific supplier.

Helios puts the rack configuration ahead of the building layout. A room designed around familiar rack widths or an older GPU load allowance can miss the power connections, liquid interfaces and access space required by the selected system. Resolve these requirements before freezing a modular facility, electrical package or outdoor cooling plant.

TL;DR

• AMD announced Helios rack-scale solutions entering production in July 2026. Product availability and delivery dates still depend on the chosen OEM system.
• The published architecture combines 72 MI455X GPUs, 18 EPYC “Venice” CPUs and Pensando networking in an Open Rack Wide arrangement.
• Use the up-to-246 kW figure from the AMD/Schneider reference design to understand the planning scale. Obtain the selected OEM’s actual load schedule before sizing infrastructure.
• Confirm loaded dimensions, mass, movement routes and maintenance space; a wider rack cannot be converted into a rack count by a simple width ratio.
• Request liquid and residual-air heat loads separately, together with coolant, pressure, temperature and control requirements.
• Keep compute-rack load, network/support IT and total facility input as separate quantities.

What Helios is, and what is available in 2026

AMD’s Helios product information describes a reference design for OEM and ODM implementation. It combines the MI455X GPU architecture, Venice processors and Pensando networking with centralized rack power and liquid-cooling interfaces. Open Rack Wide geometry is part of that design approach.

AMD’s 23 July 2026 announcement states that Helios rack-scale solutions are now in production and specifies 72 MI455X GPUs and 18 Venice CPUs. This is a more current status statement than the product-page FAQ’s earlier expectation of volume availability in the second half of 2026.

Production status does not settle which configuration you can purchase, its regional electrical arrangement or its committed delivery date. Obtain those details from the OEM quotation. Keep a versioned equipment schedule so that a later change in power shelves, network equipment or service requirements triggers a facility review.

For the broader process of translating a workload into infrastructure, see our GPU data center planning guide.

How much power does a Helios rack need?

The AMD/Schneider Electric reference-design announcement describes support for racks up to 246 kW. It reports an ANSI-based design for the US market, with an IEC extension planned. That figure and jurisdiction belong to this reference design; neither is a universal specification for every OEM Helios deployment.

A facility engineer needs several quantities that a single headline kilowatt figure cannot supply:

• Maximum rack input: the supplier’s specified electrical input for the selected equipment and operating limits.
• Expected operating load: a workload-based estimate used for energy and staged-capacity planning.
• Supply behavior: voltage, frequency, phases, power quality, connection arrangement and response to loss of a feed.
• Dynamic behavior: permitted power steps, startup sequence and any controls that limit or coordinate demand.
• Support load: networking, storage and other equipment outside the compute-rack schedule.

Ask the OEM to identify the measurement boundary and any included conversion losses. Size protective devices, conductors and supply equipment against the agreed design case and applicable electrical requirements. An expected average workload does not replace the equipment’s supported maximum input or its fault and maintenance cases.

Our data center power distribution guide follows the path from site connection to rack. A Helios project needs that path documented using the chosen rack’s actual interfaces.

A four-rack calculation establishes scale, not site capacity

Suppose an early study uses four racks at the reference-design envelope of 246 kW each:

Compute-rack input allowance = 4 × 246 kW = 984 kW.

This is an illustrative calculation, not a published ModulEdge configuration or a confirmed OEM load schedule. It excludes separate network and storage equipment, cooling auxiliaries, UPS losses and other facility loads. At six racks, the same assumed allowance becomes 1,476 kW. Changing the actual rack allowance changes the result proportionally.

Do not use 984 kW as the required utility capacity. Add the separately defined loads, model conversion and cooling losses at the site conditions, and check the agreed redundancy and maintenance cases. For phased deployment, distinguish equipment installed for the first phase from infrastructure reserved for a later phase.

Rack width changes layout and service access

Helios uses an Open Rack Wide approach. The purchasing package needs dimensional drawings for the actual assembled system, including its installed and transport states. Ask for connection locations, protrusions, doors, service zones, lifting provisions and loaded support reactions.

Width alone is an incomplete layout input. A double-width arrangement does not establish that one Helios rack can replace two standard positions in a previously drawn room. The maintenance route, liquid manifold, busbar access, cable bend space and adjacent equipment can determine the usable layout.

Check the full delivery route before fixing the building: road and unloading access, transport limits, module openings, floor transitions and the final rack movement. The rack’s operating mass and point loads also need a structural review. Reserve a practical route for replacing the largest serviceable assembly after the facility is occupied.

For a modular building, ask the infrastructure provider and OEM to review the same layout revision. Put dimensions, clearance requirements and equipment mass in the interface schedule. A visually plausible arrangement is insufficient if a rack cannot be installed or maintained through the proposed openings.

Liquid cooling leaves a separate air-side task

AMD’s reference architecture includes liquid-cooling manifolds and quick connections. The selected OEM must provide the thermal duties and operating limits for the complete rack. These include the heat transferred into liquid, the heat released into the room and the conditions under which those values apply.

Avoid assigning a generic percentage to Helios liquid heat capture. A facility design needs the actual thermal split for its equipment, including power and networking components. The room still needs a means of removing the residual air heat, and the site needs a path for rejecting the combined heat outdoors.

For the liquid interface, request the coolant specification, supply and return temperatures, required flow, pressure limits, pressure-drop curve, filtration and fluid-quality requirements. Identify who supplies the CDU, pumps, manifolds and control logic. Record whether limits apply at the facility connection, CDU or rack inlet; those are different measurement points.

The CDU sizing guide explains thermal and hydraulic selection. For Helios, feed it the OEM’s data instead of transferring a flow rate or temperature schedule from another GPU platform.

Outdoor cooling must be evaluated against the local design conditions and the required operation during equipment maintenance. A high allowable coolant temperature can create useful design options, but it does not by itself promise compressor-free operation, a fixed PUE or zero water consumption. The complete heat-rejection arrangement determines those outcomes. See our guides to liquid-cooling architecture and cooling in extreme heat.

Close the interfaces before issuing the facility specification

Use one interface schedule shared by the OEM, facility provider and site engineer. It should identify the responsible party, document revision and unresolved information for each connection.

Several items may remain provisional early in procurement. Mark them as such and identify which design decisions depend on them. For example, an unresolved rack connector can leave a pipework connection open without preventing a site survey. An unresolved maximum rack input can prevent final electrical sizing.

Commission the assembled system under the agreed operating cases

Acceptance should demonstrate the selected rack and infrastructure together. Record rack input, component temperatures, coolant conditions, residual-air conditions and useful workload during the agreed load test. A cooling-plant test without the relevant equipment behavior leaves a different question unanswered.

Agree how to test loss of a power feed, loss of a cooling component and planned maintenance. The procedure must be approved by the equipment suppliers; do not assume the rack will continue at full load or that every fault requires the same response. Record alarm destinations, control ownership and the conditions for restart.

Service access also deserves a witnessed check. Demonstrate a representative tray or component replacement using the planned lifting aids, disconnect procedure and work area. A commissioning record should show which operating cases were tested and which still rely on supplier documentation or a later site test.

Where ModulEdge fits a Helios deployment

ModulEdge’s MDC2000 is a configurable platform for NVIDIA GB300/NVL72 and AMD Helios-class infrastructure. This gives a Helios buyer a modular facility route in which rack layout, electrical distribution, cooling and service access can be developed around the selected OEM system.

The configuration needs its own rack population and electrical and thermal schedules. Published GB300 reference values do not transfer to Helios. The wider equipment, connection arrangement and access requirements shape the resulting layout and IT capacity. Platform configurability is also distinct from a claim of OEM certification or an already completed Helios installation.

Bring the OEM drawings and interface data into the discussion early. Request a ModulEdge configuration brief identifying the proposed OEM rack, supported interfaces and outstanding engineering inputs. ModulEdge can then assess the module arrangement and facility connections against that hardware requirement, while unresolved inputs remain visible in the design brief.

Frequently asked questions

Is Helios a complete AMD product that I can buy directly?

AMD describes Helios as a reference design for OEM and ODM implementations. Purchase and support terms follow the system offered by the selected supplier.

Is Helios still only a future roadmap item?

AMD’s July 2026 announcement states that Helios rack-scale solutions are in production. Confirm availability and delivery for the exact OEM configuration and market you intend to buy in.

Does every Helios rack draw 246 kW?

No universal rating follows from that figure. The AMD/Schneider reference design supports racks up to 246 kW; obtain the selected OEM’s input schedule for your facility design.

Can I use a GB300 rack layout for Helios?

Only after reviewing the actual dimensions, service clearances and interfaces. Do not transfer rack counts, support allowances or electrical and thermal totals between platforms.

Does liquid cooling remove the need for room cooling?

The facility still needs to handle the selected equipment’s residual-air heat. Request separate liquid and air duties and check their operating conditions.

Can MDC2000 accommodate Helios-class infrastructure?

It can be configured for it. Rack quantity, internal layout, resulting IT capacity and facility interfaces depend on the selected configuration and engineering review.