How to source an ODM GPU server for a data center

Decide what you are buying

An ODM builds a platform; an OEM may brand and support it; an integrator specifies, assembles and validates a customer configuration. Those roles can overlap, so compare the actual deliverable rather than the seller's label. TORmem's ODM and OEM sourcing page explains the roles; this guide turns them into a purchase checklist.

For the manufacturing-level boundary, see L10 versus L11. Ask for an itemized scope because vendors use these labels differently.

Write the workload and site brief first

A quote becomes comparable when every bidder sizes the same job. State whether the cluster is for training, inference, rendering, simulation or resale; name the required accelerator class, memory capacity, storage profile, fabric bandwidth and target deployment date.

• Workload: model or job type, target throughput, dataset size, memory bottleneck and acceptable downtime.
• Compute: GPU count and form factor, CPU sockets, DIMM population, local storage, PCIe lane needs and management interface.
• Network: backend fabric, NIC or DPU, optics, front-end connectivity and any rail or rack topology requirement.
• Site: available voltage, feed count, power per rack, cooling method, rack depth, floor loading and receiving constraints.
• Commercial: quantity, phase-in schedule, warranty location, spares, acceptance location and named support owner.

A platform that physically fits eight GPUs is not automatically validated for the eight accelerators you intend to install. ASRock Rack publishes GPU compatibility by system; use the exact platform and accelerator revision, then obtain the integrator's thermal and cabling evidence.

Make competing quotes comparable

Require an exact manufacturer part number and revision in every quote. A similar suffix may change chassis depth, cooling or supported accelerator combinations. Compare the complete BOM and assumptions side by side before using any headline price.

The AMD EPYC Turin GPU server comparison illustrates how form factor, GPU count and market references can be presented without treating a barebone market observation as a configured-system offer.

Agree on the test evidence before ordering

Write down the pass criteria, who runs each test, where it is run and what evidence will be handed over. A power-on photo is not an acceptance report.

1. Verify serial numbers and BOM against the purchase order, including firmware and GPU revisions.
2. Run node health, memory, storage and accelerator diagnostics; record errors and remediation.
3. Check NIC placement, link speed, optics and the fabric topology against the cable plan.
4. Run a representative workload or agreed synthetic test under sustained load, with power and thermal telemetry.
5. For a rack, verify power-feed loss behavior, management reachability, cabling, inter-node collectives and handover documents.

The L10 guide explains node integration, and the L11 guide covers rack build scope. Use the GPU cluster acceptance guide to turn both into a signed test matrix and handover package. NVIDIA also treats rack power, cooling and network placement as linked planning inputs for its systems.

Choose the next fork

• ODM/OEM sourcing: choose the sourcing and integration counterparty.
• GPU server family: compare concrete platforms before an RFQ.
• L10 versus L11: settle where supplier responsibility ends.
• GPU rack power: check whether the site can receive the proposed rack.
• Procurement: send a single brief covering compute, memory, fabric and site needs.

Frequently asked questions

What is an ODM GPU server?

It is a GPU server platform produced by an original design manufacturer. The platform may be sold as a barebone or configured by an OEM or integrator. Ask which company owns the chassis design, the finished BOM, the warranty and the deployment test; the exact board and revision matter more than the sales label.

What is the difference between a barebone GPU server and a configured server?

A barebone is the platform baseline and may omit CPUs, memory, GPUs, storage, add-in networking and services. A configured server has an agreed component list and should be assembled, flashed and tested against that list. Request separate prices for excluded components, integration, freight, support and acceptance so the comparison reflects the system you will operate.

Should we buy directly from an ODM, an OEM or an integrator?

Choose according to the deliverable and support boundary you need. Direct ODM supply can fit buyers with their own configuration and validation teams; an OEM or integrator can own more of the configured server, warranty and rack acceptance. Put the responsibility matrix in the RFQ: who validates GPUs, handles field failures, maintains firmware and replaces a failed part?

How do we compare two GPU server quotes?

Normalize the exact platform revision, included components, GPU power and thermal limits, firmware, tests, warranty, logistics and delivery dates. A lower barebone price is not a lower deployed-system price unless those scopes match. Ask both bidders for a line-by-line BOM and mark every assumption or exclusion before comparing total delivered cost.

Does an eight-GPU chassis support any eight GPUs?

No. Mechanical space is one condition. The exact accelerator also needs compatible power cables, PCIe layout, cooling, firmware and a validated operating envelope at the site temperature. Ask for the platform maker’s compatibility list, the integrator’s exact GPU configuration, and sustained-load power and thermal results.

What should an ODM GPU server RFQ include?

Include the workload, accelerator model and quantity, CPU and memory target, local storage, NIC and fabric, site power and cooling, rack scope, quantity, delivery schedule, acceptance tests, support location and warranty expectations. Name the target service level and ask the supplier to flag requirements it cannot verify rather than substituting a nearby platform.

What tests should be completed before shipment?

At minimum, confirm the BOM and firmware, run node health and sustained load tests, capture power and thermal behavior, and validate the agreed network links. A rack order should also include cabling, feed-failure behavior, inter-node tests and handover evidence. Agree the pass threshold, test duration and remediation owner in writing before assembly starts.

When should we buy L11 rack integration instead of L10 servers?

Choose L11 when you need a supplier to assemble and validate the rack as a unit before shipment. L10 suits a buyer who can rack, cable, power and test the servers at the destination. Define the boundary explicitly because vendors use the level names differently, and confirm whether site-side power, cooling and final cluster acceptance are included.