---
title: "AMD Helios: Identify the Configuration and Site Inputs"
description: "Leviathan Systems — Technical content Published 2026-10-17; updated 2026-10-19."
canonical: "https://www.leviathansystems.co/articles/amd-helios-configuration-site-inputs"
last-updated: "2026-10-19"
---

# AMD Helios: Identify the Configuration and Site Inputs

Leviathan Systems — Technical content
Published 2026-10-17; updated 2026-10-19.

AMD announced Helios in production on July 23, 2026 as a rack-scale reference platform combining 72 MI455X GPUs and 18 Venice CPUs. For physical planning, identify the selected OEM implementation and obtain its actual dimensions, handling, facility-interface and cabling documents, because the reference does not establish supplier stock, shipment dates or site acceptance.

## What does the Helios name establish?

AMD’s current FAQ describes Helios as an open reference blueprint that OEM and ODM partners can use for their own branded systems. AMD’s dated July 23, 2026 launch announcement states that Helios rackscale solutions are in production. Read these statements together: there is a launched platform and partner ecosystem, but an architecture name is not a purchase-ready specification for an individual site.

The first planning record should name the supplier, complete system SKU, configuration revision, quoted quantity, delivery format and supported installation package. Record the supplier’s availability, regional support and warranty position separately. A published launch date does not confirm a buyer’s shipment date or approve the facility.

This guide helps an operator, colocation buyer or project coordinator request that package. It does not compare inference benchmark results, authorize internal rack modifications or provide instructions for electrical or liquid-cooling work.

References: [A07](https://www.amd.com/en/products/rackscale-solutions/helios.html), [A10](https://ir.amd.com/news-events/press-releases/detail/1294/aai-2026-amd-delivers-full-stack-compute-for-the-agentic-ai-era)

## Keep accelerator, rack and facility reference facts separate

The current AMD blueprint identifies a 72-MI455X rack with 18 Venice CPUs, 18 compute trays and six switch trays. It uses Open Rack Wide and describes a liquid-cooled architecture. The table below identifies what each source can establish and what it leaves for the purchased system.

MI455X is a different generation and module family from MI350X or MI355X. A buyer should not reuse an MI350 OAM server elevation or power assumption as a Helios rack plan. The OEM’s engineering package is the authority for the actual cabinet, internal assemblies and external interfaces.

References: [A07](https://www.amd.com/en/products/rackscale-solutions/helios.html), [A08](https://www.amd.com/en/products/accelerators/instinct/mi400/mi455x.html), [A09](https://www.amd.com/content/dam/amd/en/documents/products/accelerators/instinct/amd-instinct-helios-blueprint-brochure.pdf)

Helios reference facts and their limits
| Level | Source-backed reference | What it does not decide | Source |
| --- | --- | --- | --- |
| MI455X accelerator | 432 GB HBM4; Enhanced Accelerator Module; dedicated product page | Complete rack consumption, cabinet dimensions or facility inlet | A08 |
| Helios reference compute tray | Four MI455X GPUs and one Venice CPU | Supplier assembly access, tray handling or field installation method | A09 |
| Helios reference rack | 72 GPUs, 18 CPUs and 18 compute trays; rounded up to 31 TB HBM4 | Actual populated weight, footprint, selected CPU SKU or cable count | A09 |
| Partner facility reference | Schneider/AMD design supports an up to 246 kW rack envelope | Universal Helios consumption or a site-approved electrical design | O06 |
These are attributed reference facts checked October 14, 2026, not construction instructions. The supplier and facility design owners must reconcile the selected implementation.

## Count reference units before counting project items

The Helios blueprint describes nested units, not 72 independent rack servers. Its compute tray combines four MI455X GPUs with one Venice CPU; the reference rack has 18 such compute trays and six switch trays. Accelerator, compute-tray, switch-tray and rack counts answer different inventory questions. A drawing that identifies a tray position does not necessarily identify a separately shipped item or a field-removable assembly for the appointed installer.

Fictional EX-HL109 compares a request for 144 GPUs with a worksheet that explicitly assumes two unchanged copies of the published reference population. Under that assumption, 2 × 72 = 144 GPUs, 2 × 18 = 36 compute trays and 36 CPUs, and 2 × 6 = 12 reference switch trays. These are reference-population arithmetic, not an OEM bill of materials, a shipment count or permission to assemble the equipment. In particular, 144 GPUs does not mean 144 server chassis, 144 external cables or 144 independently installable tasks.

If an OEM offer states only the GPU quantity and architecture name, the compute-tray and switch population are still to be confirmed for that implementation. Preserve the distinction between a quantity normalized from the reference and a quantity explicitly identified by the supplier. The selected equipment and issued interface map must determine the actual external connection population; neither the GPU count nor the reference scale-up bandwidth supplies it.

References: [A09](https://www.amd.com/content/dam/amd/en/documents/products/accelerators/instinct/amd-instinct-helios-blueprint-brochure.pdf)

EX-HL109: two stipulated reference populations, not a purchased configuration
| Unit | One published reference rack | Two-copy worksheet | What remains unestablished | Source |
| --- | --- | --- | --- | --- |
| GPU | 72 MI455X GPUs | 144 GPUs | Actual supplier SKU, population and shipped condition | Planning analysis |
| Compute tray and CPU | 18 compute trays; one CPU per tray | 36 compute trays; 36 CPUs | Exact CPU SKU, tray configuration and supplier integration boundary | Planning analysis |
| Switch tray | 6 reference switch trays | 12 reference switch trays | The selected OEM fabric equipment, internal assemblies and exposed interfaces | Planning analysis |
| Rack-scale reference | 1 ORW reference rack | 2 ORW reference racks | Actual enclosures, shipping units, footprint, handling and site applicability | Planning analysis |
Original population example. Two copies are a stated arithmetic assumption, not an order, supported cluster topology, supplier commitment or approved physical work package.

## A source can be current and still contain an unresolved difference

Keep the source title, revision, publication/update date where exposed and the date checked with every important specification. The Helios page currently presents two different per-GPU memory-bandwidth figures: 23.3 TB/s in its highlights and 19.6 TB/s in a compute-tray paragraph. The dedicated MI455X product page reports 23.3 TB/s. A buyer should request the supplier’s applicable revision instead of combining the larger number with unrelated rack or benchmark figures.

The availability FAQ still describes volume deployments as expected in 2H 2026, while AMD’s dated July 2026 announcement says its rackscale solutions are in production. The production statement can be attributed to its date; it still does not replace the OEM’s order availability. This guide avoids a bandwidth ranking or performance-per-dollar conclusion, because those comparisons require matched configurations and workload assumptions.

References: [A07](https://www.amd.com/en/products/rackscale-solutions/helios.html), [A08](https://www.amd.com/en/products/accelerators/instinct/mi400/mi455x.html), [A10](https://ir.amd.com/news-events/press-releases/detail/1294/aai-2026-amd-delivers-full-stack-compute-for-the-agentic-ai-era)

## Separate the ORW interface reference from the OEM implementation

AMD’s FAQ describes Open Rack Wide as double-wide, distinguishing it from earlier single-wide OCP racks. The OCP contribution index separately lists an ORW Base Specification and an ORW Meta Design Specification, both V1.0.1 in its 2026 entries. A common “Open Rack” label therefore does not identify the same document, revision or enclosure. The ORW name is not a conversion rule for an existing ORv3 row, a handling checklist or an assurance that a proposed delivery fits its route.

A concrete OEM distinction also appears in AMD’s dated December 2, 2025 HPE collaboration announcement: it describes an HPE Juniper scale-up switch developed with Broadcom for Helios. That is a named implementation announcement, not the part list of every Helios-based rack or evidence of an available customer order. It demonstrates why the AMD reference’s switch-tray description cannot silently become an OEM-specific equipment inventory. Record the applicable supplier model and interface documentation before comparing the reference with the offer.

For the buyer, the useful comparison is between the reference architecture and the actual supplied implementation. Keep rack-standard identity, the OEM’s populated equipment and the supplier’s installation/service information distinct. A shared open-standard reference can support that comparison; it does not settle the selected electrical, liquid, structural, access or handling requirements. Those requirements still need the actual equipment package and the responsible qualified parties. The existing site-interface and project-package sections identify the inputs to request; this comparison supplies no lift, movement, internal-assembly or operating method.

References: [A07](https://www.amd.com/en/products/rackscale-solutions/helios.html), [A11](https://newsroom.amd.com/news/amd-and-hpe-expand-collaboration-to-advance-open-r/), [O07](https://www.opencompute.org/contribution-family/openrack/)

Identify which source is answering the buyer’s question
| Source level | Narrow source-backed example | Buyer decision still required | Source |
| --- | --- | --- | --- |
| Open rack documentation | OCP lists ORW Base and ORW Meta Design as distinct V1.0.1 entries | Which standard and revision does the selected supplier identify? No compliance or fit is established by the title. | Planning analysis |
| AMD architecture reference | The blueprint’s 18 compute trays, 6 switch trays and 72 GPUs | Which reference elements and quantities apply to the actual supplier configuration? | Planning analysis |
| Named OEM implementation announcement | AMD’s December 2025 HPE announcement names HPE Juniper scale-up switching with Broadcom | What equipment and external interfaces are actually in the selected offer? An announcement is not its current BOM or delivery confirmation. | Planning analysis |
| Purchased equipment and installation package | The supplier’s exact system identity, populated configuration and applicable documents must be obtained | Actual dimensions, delivery/handling, power/cooling interfaces, task access and qualified site acceptance remain configuration-specific. | Planning analysis |
Document and product identity comparison only. No ORW geometry, standards certification, OEM relationship, inventory, compatibility, handling or engineering approval is inferred.

## Request interfaces, not a universal rack power number

A GPU power specification, a rack consumption figure and a facility reference-design envelope have different meanings. The Schneider Electric and AMD July 2026 reference design describes support for up to 246 kW rack density. That is a design envelope for the published reference, not a verified load or input requirement for every OEM rack sold under the Helios architecture.

Obtain the chosen equipment’s full input, connection, redundancy, operating and heat-rejection requirements. Qualified customer-appointed electrical and mechanical owners must confirm the corresponding site design. Record any residual airflow and service-access requirements alongside the liquid-system interfaces. Exact fluid, flow, pressure and connection requirements need the supplier’s documentation and specialist disposition; they are not established by the reference headline.

Open Rack Wide is also a design standard, not an automatic fit approval for an existing cabinet row. Obtain exact supplier dimensions, shipping and populated weight, footprint, service access and receiving/placement requirements. The responsible site and structural/handling specialists must approve the actual conditions.

References: [A07](https://www.amd.com/en/products/rackscale-solutions/helios.html), [A09](https://www.amd.com/content/dam/amd/en/documents/products/accelerators/instinct/amd-instinct-helios-blueprint-brochure.pdf), [O06](https://www.se.com/uk/en/about-us/newsroom/news/press-releases/Schneider-Electric-and-AMD-Release-First-Helios-Platform-Reference-Design-to-Accelerate-AI-Factory-Deployment-6a62298cfdd18b2e110350f6/)

## Build the external cable schedule from the installed equipment

The blueprint distinguishes scale-up connectivity inside the rack from Ethernet scale-out between systems. It describes UALink over Ethernet for the rack fabric and Pensando Vulcano networking for scale-out. That helps organize the questions, but an aggregate bandwidth figure does not identify physical port locations, connector quantity, optic part numbers or the network topology purchased for a project.

Request the OEM’s approved interface drawings and the customer network owner’s port map. Keep scale-up cartridge work, external cluster connections, management and storage paths separate. Agree which internal assemblies are factory-integrated and which physical connections the appointed installer is permitted to access. Do not translate “72 GPUs” into 72 external cables without an actual endpoint list.

A point-to-point schedule can then identify each device and port, approved connector and cable/optic, route, reach, label and responsible reviewer. The physical installer’s handoff should record agreed checks and discrepancies. Configuration, software diagnostics, workload validation and any formal optical acceptance remain separately assigned tasks.

References: [A09](https://www.amd.com/content/dam/amd/en/documents/products/accelerators/instinct/amd-instinct-helios-blueprint-brochure.pdf)

## The project-input packet to request from an OEM or integrator

Use this checklist before requesting an installation price or committing a site date. Mark each input as supplied, under review or missing; record the owner of each missing item. Reconcile changes into the same package rather than letting a sales datasheet, rack layout and cable schedule describe different configurations.

- OEM/ODM supplier, full rack SKU, configuration revision, quoted equipment list and quantity.
- Reference generation and exact accelerator/CPU/networking variants.
- Confirmed order availability, delivery format, factory integration boundaries, warranty and support contacts.
- Rack/tray dimensions, shipping and populated weight, service clearances and approved handling/placement requirements.
- Complete equipment power/input/redundancy specification and the electrical owner’s site acceptance.
- Cooling/heat-rejection and airflow interface package with the appointed specialist’s readiness disposition.
- Installed scale-out, management and storage equipment inventory with approved port and topology drawings.
- Selected cable/optic BOM, routes, approved reach, labels and point-to-point schedule.
- Permitted physical access/work boundaries and named owners for internal assemblies and specialist interfaces.
- Project-specific physical verification and documentation requirements, with separate system acceptance owner.

## A responsibility matrix for the first scope meeting

The roles below are a starting point for a buyer discussion. Assign actual named owners and contract boundaries for the chosen project. A physical installation proposal should identify its prerequisites and deliverables without absorbing facility construction or full system acceptance by implication.

Illustrative Helios scope and dependency matrix
| Question or task | Owner to appoint | Reviewable output | Source |
| --- | --- | --- | --- |
| Which Helios-based system is supplied? | Customer equipment owner + OEM/ODM | Exact BOM, revision and supported installation package | Planning analysis |
| Can the site accept its input, cooling and physical envelope? | Facility design owners and qualified specialists | Documented site/interface acceptance for that configuration | Planning analysis |
| How does equipment move and sit at the site? | Site-authorized handling/placement and structural owners | Approved handling conditions, route and placement scope | Planning analysis |
| What external physical connections are required? | OEM interface owner + customer network designer | Approved endpoint map and cable/optic list | Planning analysis |
| Who mounts permitted equipment and routes/dresses/labels cables? | Named physical installer within the agreed scope | Installation plan and point-to-point schedule | Planning analysis |
| Who owns liquid-side and electrical work? | Appointed mechanical/electrical/OEM specialists | Separate authorized procedures and acceptance evidence | Planning analysis |
| Who configures and accepts the running cluster? | Customer IT, OEM and software/workload owners | Separate configuration, diagnostic and workload plan | Planning analysis |
| What proves the physical work is complete? | Customer physical reviewer + installer | Agreed inspection, labels, exceptions and handoff documentation | Planning analysis |

## Example: turn a rack headline into a defined planning task

A buyer has a Helios-based proposal and a planned colocation position. The missing information is not another headline about GPU counts. It is the supplier configuration, rack and delivery envelope, approved facility interfaces and actual network endpoint schedule. The buyer requests those records and assigns an owner to each acceptance decision.

Once supplied, physical planning can reconcile rack positions, equipment identification, cable routes, labels and scope interfaces. If the supplier changes the networking cartridge, external adapter or rack input configuration, the affected documents and approvals must be updated before work on that interface proceeds. This is an illustrative buyer workflow, not a Leviathan customer case study.

## Use the planning package for an accurate service discussion

Leviathan can discuss physical infrastructure planning, layouts, routing coordination, point-to-point cable schedules, labeling architecture, installation planning and agreed documentation. CAD production resources, supported software/file formats and the engagement’s deliverables must be confirmed before a software-specific promise is made. The service does not establish licensed engineering authority, stamped drawings or structural/electrical approval.

Leviathan also offers equipment receiving and preparation, rack assembly, structured cabling, physical deployment and agreed hardware burn-in and QA. A Helios reference design does not establish an available OEM order, an authorized internal assembly task or a qualified site. Confirm the actual supplier configuration and permitted physical work before defining the installation package. Identify the equipment/material supplier, approved configuration, quantities and delivery responsibilities before agreeing the physical work.

For hardware burn-in, define the selected equipment, OEM or customer method, conditions and duration, tools and personnel, recorded results, exceptions, retest decisions and acceptance criteria. That scope does not include a promise of complete fabric or workload acceptance. Physical relocation and refresh can be discussed against the actual inventory, origin and destination, custody arrangements and site requirements.

Liquid-side work and facility electrical construction or energization remain outside the standard scope. BIOS, firmware, IP configuration, OS imaging, software administration and full software commissioning are not standard services. Share the selected supplier, project location, equipment stage and available interface documents to discuss the scope, requirements and availability.

## Common planning questions

### Is Helios an AMD rack I can buy directly?

AMD’s FAQ describes Helios as a reference design for branded OEM/ODM systems. The buyer must identify the actual supplier, system SKU, revision, availability and support position.

### Is Helios still only a roadmap item?

AMD’s July 23, 2026 launch announcement states Helios rackscale solutions are in production. That dated platform statement does not confirm stock, shipment or site approval for a particular OEM order.

### Does every Helios rack require 246 kW?

The 246 kW figure is an up-to envelope in a particular Schneider/AMD facility reference design. Obtain the chosen OEM equipment’s full input and the facility owner’s approved design; do not use it as universal consumption.

### Can a MI350 server planning package be reused unchanged?

MI350-series OAM servers and the MI455X Helios rack architecture are different configurations. Existing labels or document formats may be reused after reconciliation, but equipment, interface and cable facts need the chosen Helios OEM package.

### Does the rack bandwidth figure tell me how many cables to install?

No. Use installed network equipment, physical interface drawings and the approved endpoint port map. Aggregate bandwidth is not a connector, optic or cable inventory.

### Can Leviathan prepare layouts and cable schedules?

Leviathan can discuss project-defined physical planning, routing, point-to-point schedules, labels and documentation. Production resources, software/file formats and review deliverables must be confirmed. Licensed/stamped engineering, liquid-side work and electrical construction are outside the offered scope.

### Does a 144-GPU request define two purchased Helios racks?

Only the fictional EX-HL109 worksheet stipulates two unchanged copies of the published 72-GPU reference. Its arithmetic gives 36 compute trays, 36 CPUs and 12 reference switch trays. A real offer still needs the exact supplier configuration and integration/delivery format; a GPU quantity alone does not identify enclosures, shipping units, external cables or installation tasks.

### Does the ORW name make the reference a complete OEM installation specification?

No. Identify the exact standard and revision, the selected OEM implementation and its own installation/interface package. OCP lists distinct ORW Base and Meta Design entries, and the named AMD/HPE announcement illustrates implementation-specific switching. Neither the reference nor that announcement supplies the actual equipment’s handling, electrical, liquid, structural, service-access or site acceptance decisions.

## Primary sources

- [A07 — AMD Helios reference design and FAQ](https://www.amd.com/en/products/rackscale-solutions/helios.html). Checked 2026-10-14. Published reference; exact purchased OEM configuration and revision control. Availability statements do not establish a customer delivery date.
- [A08 — AMD Instinct MI455X product specifications](https://www.amd.com/en/products/accelerators/instinct/mi400/mi455x.html). Checked 2026-10-14. Published reference; exact purchased OEM configuration and revision control. Availability statements do not establish a customer delivery date.
- [A09 — AMD Helios blueprint brochure](https://www.amd.com/content/dam/amd/en/documents/products/accelerators/instinct/amd-instinct-helios-blueprint-brochure.pdf). Checked 2026-10-14. Published reference; exact purchased OEM configuration and revision control. Availability statements do not establish a customer delivery date.
- [A10 — AMD July 23 2026 launch and Helios production announcement](https://ir.amd.com/news-events/press-releases/detail/1294/aai-2026-amd-delivers-full-stack-compute-for-the-agentic-ai-era). Checked 2026-10-14. Published reference; exact purchased OEM configuration and revision control. Availability statements do not establish a customer delivery date.
- [O06 — Schneider Electric UK / AMD July 23 2026 Helios facility reference design announcement](https://www.se.com/uk/en/about-us/newsroom/news/press-releases/Schneider-Electric-and-AMD-Release-First-Helios-Platform-Reference-Design-to-Accelerate-AI-Factory-Deployment-6a62298cfdd18b2e110350f6/). Checked 2026-10-17. Official UK edition of the July 23 announcement. The up-to 246 kW figure describes this facility reference design, not universal rack consumption or an approved customer installation. Confirm the purchased OEM configuration and qualified facility design.
- [A11 — AMD December 2, 2025: HPE Helios implementation collaboration](https://newsroom.amd.com/news/amd-and-hpe-expand-collaboration-to-advance-open-r/). Checked 2026-10-19. Only the dated announcement’s named HPE Juniper scale-up switch/Broadcom collaboration illustrates an implementation-specific element. No present product availability, shipped BOM, performance, procurement, OEM relationship or site applicability is inferred.
- [O07 — OCP contribution index: Open Rack Wide Base and Meta Design entries](https://www.opencompute.org/contribution-family/openrack/). Checked 2026-10-19. Only distinct Base and Meta Design V1.0.1 identities in the 2026 index are used. No technical dimensions, standards compliance, design/CAD rights, handling method or supplier implementation is established.

## Related Leviathan resources

- [Engineering and physical infrastructure planning](/services/engineering-design)
- [Colocation installation](/services/colocation-installation)
- [Rack-and-stack installation](/services/rack-and-stack)
- [Structured cabling](/services/structured-cabling)
- [Equipment receiving and rack assembly](/services/gpu-assembly)
- [AMD MI350X and MI355X: Physical Installation Readiness](/articles/amd-mi350-physical-installation-readiness)
- [ESD-Protective Packaging: Identify the Functions the Component Transfer Needs](/articles/esd-packaging-charge-dissipation-shielding-transfer)
- [Data Center Power Capacity: Read the Megawatts Behind a Site Claim](/articles/data-center-power-capacity-claims)

## Discuss the physical planning inputs

Share the selected OEM system, project location and available interface documents. Our team will follow up to discuss the scope, requirements and availability.

[Discuss the scope](/contact?service=engineering-design&from=/articles/amd-helios-configuration-site-inputs)