AMD MI350X and MI355X: Physical Installation Readiness

Which MI350-series system is actually arriving?

A purchase line that says “MI350 series” is insufficient for a rack elevation or cable schedule. Ask for the server model, hardware revision, accelerator SKU, CPU and memory configuration, installed network adapters, cooling variant and delivery format. Record whether the equipment arrives as an integrated server, a rack assembly or components that the OEM explicitly permits the appointed installer to mount.

This article addresses the physical planning package for MI350X or MI355X equipment, identified in the project’s approved equipment record. It does not select a GPU for a workload, establish a certified upgrade path or prescribe a cooling or electrical installation procedure. The selected OEM manual and facility requirements take precedence over a family-level comparison.

References: [A03][A04][A05][A06]

What changes between MI350X and MI355X?

Both current AMD product pages list 288 GB HBM3E and 8 TB/s memory bandwidth per GPU. The product pages label 1,000 W for MI350X and 1,400 W for MI355X as Typical Board Power (TBP). The platform brochures separately label their respective per-module figures Maximum TBP; keep each label with its source rather than treating them as one unlabeled system-power value. Those figures identify an accelerator reference; they are not the server’s facility input or a safe rack-density limit. CPUs, memory, storage, fans, power-conversion losses and the selected operating configuration also matter.

AMD’s platform brochures describe eight OAM accelerators on a UBB 2.0 baseboard and publish a rounded 2.3 TB aggregate GPU-memory total. That total does not establish a workload fit guarantee, a whole-rack capacity or a physical NIC inventory. Reconcile the exact platform revision with the supplier instead of treating a shared family name as approval for a module swap.

References: [A03][A04][A05][A06]

Confirm the server cooling variant before selecting a location

The server model determines the actual cooling implementation. Supermicro documents an 8U air-cooled AS-8126GS-TNMR supporting MI350X and a 4U direct-to-chip liquid-cooled AS-4126GS-NMR-LCC with MI355X. Supermicro also announced an air-cooled MI355X system in November 2025. Consequently, “MI355X requires liquid cooling in every system” is not a valid planning rule.

For an air-cooled order, obtain the supplier’s airflow, inlet conditions, exhaust arrangement and service-access requirements; the facility owner must accept that equipment in the intended location. For a liquid-cooled order, obtain the OEM interface package and the appointed mechanical provider’s readiness disposition. Connection details, fluid requirements and thermal acceptance belong to those qualified parties. Leviathan does not install, connect, disconnect, test or commission liquid-side equipment.

A power-supply rating is also different from equipment consumption. A list of redundant supplies cannot be added together and presented as the customer’s required load without the OEM’s complete power specification and the approved facility design.

References: [O01][O02][O03]

Distinguish MI355X rack occupancy from the access envelope

An MI355X accelerator name does not determine a rack position. Even after the complete server is identified, distinguish its rack-height allocation, chassis body and required access arrangement. A blank rectangle in an elevation establishes neither rail compatibility nor space for the applicable cables, liquid interfaces and service activity.

Supermicro's AS-4126GS-NMR-LCC page labels the system 4U rackmount and separately lists a chassis body 449 mm wide, 174 mm high and 895.35 mm deep. Those are different fields: 4U is not a complete three-dimensional fit statement, and chassis depth is not the mounting-post adjustment range, a connected-system envelope or an aisle requirement. The product page also lists separate packaging dimensions; do not replace an installed-body dimension with a shipment field. [O07]

The MNL-2801 model-family access page identifies CPU and GPU trays at the front and an I/O tray at the rear. This makes a front-only access assumption incomplete for the described family. It does not supply an actual clearance, an extraction path or permission to service the purchased configuration. Use the applicable OEM arrangement to assess each intended access task, including what the connected equipment and cabinet allow. No tray-removal or liquid-side procedure is reproduced here. [O08]

Fictional EX-MA155 compares three planning assertions for two proposed instances of the named 4U system. A reserved eight-U block agrees with the sum of their two 4U allocation labels. A separate 1,200 mm figure describes the cabinet's outside depth, not usable internal depth. Subtracting the published 895.35 mm chassis depth gives 304.65 mm, but that number is only a difference between unlike reference boundaries; it is not rear clearance or a cable/connector allowance. A third assertion addresses front access only and leaves the described rear I/O tray unaccounted for.

There are three comparison lines: one limited height-label agreement and two unresolved placement/access assertions. There are zero approved rack positions or service arrangements. The eight-U label agreement does not close the other two lines. The example does not locate either server at a particular U position or determine whether two units should share a rack.

For the actual placement decision, compare the cabinet's applicable internal and mounting dimensions, the selected rail/accessory arrangement, the connected equipment envelope and the intended service tasks against the same ordered configuration. Keep the liquid connections, permitted cable routing, loading/handling and facility conditions with their responsible qualified owners. A returned front-and-rear arrangement must account for the actual components and work state; a body-depth subtraction cannot stand in for it. The existing responsibility matrix and installation package still apply, including the supplier's service-personnel and liquid-cooling conditions.

References: [O07][O08]

An accelerator specification and a server installation manual answer different questions

An AMD specification identifies the accelerator. An OEM product page identifies a server offering. A model-specific manual addresses that server’s installation and use, subject to its stated applicability and personnel requirements. These documents belong together, but one cannot silently substitute for another.

For a concrete example, AMD’s MI355X page identifies an OAM accelerator with 288 GB HBM3E and a 1,400 W Typical Board Power. Supermicro separately lists the AS-4126GS-NMR-LCC as a 4U liquid-cooled server supporting MI355X and states an onsite-service requirement for liquid cooling. The component’s board-power field is not that complete server’s facility input. The server page’s service condition is a dependency to resolve with the supplier, not evidence that a particular physical contractor is eligible to perform liquid-side work.

The model’s resource center lists a user manual at revision 1.0 and a quick reference at revision 1.00 as separate resources. The online MNL-2801 preface identifies the AS-4126GS-NBR/NMR-LCC models, addresses professional integrators and technicians, and restricts installation and maintenance to certified service technicians. Reading a public manual does not establish an appointment, qualification or permission to execute its procedures.

Keep the document number, covered model suffixes, revision, publication date and retrieval date as separate fields. The online preface displays August 15, 2025 as the manual’s publication date and August 08, 2025 as that page’s update date; neither is the resource center’s revision field. Ask the supplier to identify the document and supplements applicable to the ordered configuration. A recent download date cannot resolve a wrong model or an unconfirmed option.

References: [A11][O04][O05][O06]

Resolve a mixed-model packet without inventing an installation release

In fictional example EX-AM97, an equipment line names an MI355X AS-4126GS-NMR-LCC server. The buyer receives three records: an MI355X accelerator sheet, an AMD eight-GPU platform brochure and a server manual whose cover names AS-8126GS-TNMR. The first two records provide component and platform context. The third names a different OEM server from the equipment line. None of the three establishes an installation release for the offered 4U configuration.

Record the mismatch precisely: the offered server’s applicable manual and option supplements have not been established. Return that question to the equipment supplier, with the exact offered model and configuration. Retain the other-model manual as a mismatched reference rather than copying its rail, access or handling requirements into the intended work package. There are three supplied documents, two usable reference-level records and one different-model manual; there are zero approved physical tasks in this example. This is a document-selection result, not a finding that the actual server is defective or unsupported.

If the supplier returns the correct model-family manual, that resolves one identity mismatch. It does not alone resolve installed options, the required service arrangement, cabinet suitability, access for the actual equipment, facility acceptance or permission for each party’s work. Record those decisions against their own evidence and owners. Conversely, a missing or inapplicable manual does not prove that every rail, cable or facility interface must change.

The buyer can now distinguish three different answers: the accelerator is identified; the offered server documentation needs reconciliation; actual placement and service access remain unaccepted. The retained equipment checklist and responsibility table provide the rest of the inquiry inputs. This example adds no rack location, access distance, supported conversion, liquid connection or installation method.

References: [A03][A04][A05][A06][O01][O04][O05][O06]

Translate the installed NIC inventory into a cable package

Infinity Fabric within the AMD GPU platform and the external scale-out network answer different questions. The internal interconnect does not tell the installer how many external cables to route. Use the server’s installed adapter and switch part numbers to build an endpoint list. Separate management, storage, front-end and cluster connections where the approved architecture requires them.

For each external connection, the cable schedule should identify both device names, both port labels, connector and optic/cable part numbers, quantity, approved reach, route and final label. The network design owner must approve compatibility and topology. The physical team can then reconcile pathways and routing with the permitted installation scope. A stated “400G” or “800G” speed does not uniquely identify the connector, fiber count, breakout or polarity.

Useful handoff evidence is a reconciled schedule, labels that match it, a recorded list of deviations and the specific physical verification results agreed for the project. Formal certification, optical loss testing or software fabric diagnostics need their own confirmed method, personnel, equipment and acceptance criteria.

References: [A05][A06][O01][O02]

The input checklist that makes an installation scope reviewable

Keep the package versioned: one equipment list, one rack elevation, one approved port map and one recorded set of facility dispositions. Changes to NICs, rack placement or cooling variant should be reconciled into that package before the affected physical work is released. An unresolved input should identify its owner and the decision it blocks.

• →OEM model, ordered accelerator variant, hardware revision and supported installation document.
• →Equipment quantities, delivery format, receiving constraints and an approved handling/placement plan.
• →Supplier dimensions, configured weight, compatible rack/rails and required service access.
• →Facility acceptance for the complete equipment input, cooling/airflow interfaces and structural conditions.
• →Installed server and switch NIC inventory, approved optics/cables and endpoint port map.
• →Rack elevation, pathway capacity, route, labeling scheme and point-to-point schedule.
• →List of activities permitted by the OEM and site, with named customer and specialist interfaces.
• →Agreed physical inspection, defect recording, verification and handoff deliverables.

Assign the work before comparing installation quotes

Use this table as a discussion template, then replace the owner roles with actual names and the agreed engagement boundaries. It is not a universal assignment for every site. A quote becomes more useful when the buyer can distinguish an included physical task from a dependency owned by another party.

A practical planning example

Suppose the purchase schedule identifies an eight-GPU MI350-series server, while the site plan only reserves a rack position and a generic network uplink. The next useful action is to reconcile the server’s exact variant, configured dimensions and rails, facility-approved inputs, installed NICs and switch ports. That produces a concrete cable and placement package.

If the supplier then changes the cooling variant or network adapters, the relevant facility and cable approvals must be reconciled before that equipment is installed. The GPU family name may remain the same while the physical work changes. This is an illustrative planning case, not a record of a Leviathan customer deployment.

Physical planning and the next project discussion

Leviathan offers equipment receiving and preparation, rack assembly, structured cabling, onsite physical deployment and agreed hardware burn-in and QA. Customer-furnished equipment remains a valid project model. The ordered OEM configuration, authorized work, site requirements and agreed delivery resources define the engagement; a GPU family name does not confirm stock, compatible equipment or a site-ready system.

For this installation package, bring the selected server model, quantities, supply responsibility, destination, rack and cable documents, receiving constraints and required handoff records. If hardware burn-in is requested, define the equipment, applicable OEM or customer method, duration and conditions, tools and personnel, results, exceptions, retest decisions and acceptance criteria before the work begins. Hardware QA does not establish network, ROCm or workload acceptance.

Liquid-side installation or connections, facility electrical construction or energization and licensed or stamped engineering remain outside the standard scope. BIOS, firmware, IP configuration, OS imaging, software administration and complete software commissioning are not standard offerings. The team will follow up to discuss project scope, requirements and availability.