Services

Specification and configuration

A GPU server is only as good as the configuration behind it: the firmware set has to be consistent, the NVMe laid out for the workload, the fabric cards in the right slots for the PCIe topology, and the whole thing run hard enough before delivery to shake out early failures. We own that specification on your behalf.

Most systems, and every HGX node and rack scale product, come from the manufacturer configured to the bill of materials we agree with you. We verify firmware, slot layout and the operating system and driver stack on arrival, then test. For PCIe GPU servers and workstations we can also supply our own white label configurations, assembled in San Diego from manufacturer chassis and components under Computer Modules part numbers, which gives us more control over lead time and firmware baseline on those classes. The quotation states which route applies.

For multi node orders we specify the nodes, the fabric switches and the cabling plan together, arrange racking and cabling through the manufacturer or an integration partner, and run the cluster level acceptance tests before the order ships to your site, either as racked and cabled units or as labeled components with a cabling plan, depending on how your site receives freight.

Acceptance testing

Every system is run through the same five hardware checks before delivery. You receive the results as an acceptance report with the serial numbers, firmware versions and measured figures, so the numbers in your runbook are the ones that were measured on your hardware, not the ones on a datasheet.

Check	What it proves	Tool
HW-1	Collective throughput across all GPUs in a node, and across nodes on cluster orders	nccl-tests (all_reduce and all_gather)
HW-2	Every InfiniBand or RoCE rail reaches line rate, on its own and all at once	ib_write_bw per rail
HW-3	GPUDirect RDMA works end to end and the NVMe stack reads and writes at the expected rate	GPUDirect RDMA and NVMe smoke tests
HW-4	NVLink and NVSwitch links are complete and run at full bandwidth	nvbandwidth
HW-5	The node holds a sustained power cap for 24 hours with no thermal throttling, ECC faults or link flaps	24 hour soak under load

A single failing GPU, cable or transceiver is replaced and the affected test is run again before anything leaves. The same five checks run on every order, from a single workstation to a multi rack cluster, so systems bought at different times behave the same way on day one.

Warranty options

Standard cover on servers and workstations is three years parts and labor, with advance replacement on GPUs and drives where the manufacturer supports it. You can add next business day onsite service for sites in the United States and Canada, or 4 hour onsite response in metropolitan areas where the service network supports it. Five year terms are available on most server classes and we recommend them for anything that will stay in production beyond the usual refresh cycle.

The manufacturer's warranty passes through to you in full, and Computer Modules handles the claim on your behalf so you have one contact rather than several. Cover details are stated on each quotation.

Logistics and export

We ship from San Diego to anywhere in the United States and Canada. Systems travel on pallets in the manufacturer's packaging or in shock monitored crates for racked clusters. Our customs broker prepares the commercial invoice and HS codes before the carrier collects, so Canadian shipments clear without a hold, and we can act as, or arrange, the importer of record where your organization prefers not to.

Data center GPUs need NVIDIA end user registration and a US export compliance check before shipment. We collect the end user, deployment address and end use with your quotation and run the checks while the system is being built, which is why they rarely add time. The rules are explained in plain English on the export compliance page.

Financing and leasing

Most AI hardware we sell can be leased rather than bought. We work with third party leasing partners who understand GPU residual values and can structure 24, 36 or 48 month terms, fair market value or nominal buyout at the end, and in some cases a deferred start so the first payment lands after the cluster is in production. Ask for a lease quotation alongside the purchase quotation and we will arrange both. Credit decisions sit with the leasing partner, not with us.

RMA and support

If something fails, email support with the serial number and a short description and you will receive an RMA number and instructions the same business day. GPUs, drives and power supplies are usually advance replaced; whole systems are repaired through the manufacturer's service network, coordinated by us. We keep the firmware baseline and acceptance report for every system we ship, so a replacement part is brought to the same baseline before it goes back to you.