What is L10 server integration?

Where L10 sits on the integration scale

On the L1–L12 manufacturing scale, L6 is the barebone: a motherboard integrated into its chassis and power-on tested. L7 to L9 add expansion cards and GPUs, then drives, then CPUs and memory. L10 is the first level at which the server is complete and tested as a whole system, or as Glenn Lockwood puts it, "a fully working server that can fully boot into an OS."

For the next step up, see L11 rack integration. For the trade-off between the two, see L10 vs L11.

Barebone vs L10: what a barebone price leaves out

Barebone prices are easy to find and easy to misread. A barebone is the platform: chassis, motherboard, power and cooling. On a GPU server it's usually the smaller part of the bill. Most of the cost sits in what L10 adds:

• CPUs and memory
• GPUs
• NVMe storage
• Network adapters or DPUs, and their optics
• Assembly, firmware and test labour
• Burn-in time, which moves the risk of an early-life failure from your data center to the factory

That's why TORmem shows a dated barebone market range on each server page and quotes the configured system separately. Examples: the 4U8G-TURIN2/RF+, the 4U4G-TURIN/HPR, and the full AMD EPYC Turin GPU server family.

What happens during an L10 build

1. Component install: CPUs, memory, GPUs, drives and network adapters, in the slots the platform and the workload call for.
2. Memory population: DIMMs placed to fill memory channels evenly. An unbalanced population costs memory bandwidth.
3. Firmware and BIOS baseline: BIOS, BMC, GPU, network adapter and drive firmware set to agreed versions, and BIOS options such as power profiles, NUMA and PCIe settings configured for the workload.
4. OS or image: a base operating system, or the client's own image if one is specified.
5. Records: serial numbers, MAC addresses and the final configuration captured for handover.

On liquid-cooled platforms, cold plates and quick-disconnects are usually fitted and tested at L10, ahead of rack-level manifolds and CDUs at L11.

How an L10 server is tested

L10 testing proves that each server works on its own before it joins a rack.

• Functional: the server completes POST, every component is detected at the expected speed and link width, and the BMC is reachable.
• Component diagnostics: GPU diagnostics (NVIDIA DCGM on NVIDIA platforms), memory tests, storage benchmarking with FIO, and network adapter link tests.
• Burn-in: sustained load to surface early-life failures. Third-party guidance for AI infrastructure ranges from about 72 to 168 hours.
• Final verification: re-checked after burn-in, with the results recorded.

TORmem builds to order: the burn-in duration, tools and pass criteria are agreed with each client before the build.

Why L10 tests outlive the factory

L10 tests aren't only for the factory. Rack- and cluster-level tests can disrupt healthy nodes, so on a running cluster, single-node L10 tests are often the only ones you can run on a server that has been pulled for repair. Keep the L10 test suite and its baseline results. They're how you prove a repaired node is healthy before it goes back into service.

"L10" doesn't mean one thing

The scale is a convention, and L10 is where vendors disagree most. Corning's definition in DCD treats L10 as basic mechanical rack assembly rather than a finished server. In NVIDIA's supply chain, L10 is used for a finished compute tray, and a J.P. Morgan note reported that NVIDIA would ship complete L10 trays for Vera Rubin, leaving rack integration to ODMs. NVIDIA has not confirmed this.

When a vendor quotes "L10", ask for three things: what is installed, which firmware versions, and which tests were run for how long.

When L10 is enough

• You already have racks, power, cooling and a network fabric in place.
• Your team racks, cables and validates hardware itself.
• You're adding servers to an existing cluster.
• You're deploying a handful of servers, not full racks.

If you're standing up new racks, running liquid cooling, or working to a fixed go-live date without an integration team, look at L11. The L10 vs L11 guide has a decision checklist.

Memory is decided at L10

Memory configuration is fixed when the server is built. Current server CPUs have many memory channels (AMD EPYC 9005 "Turin" has 12 DDR5 channels per socket), and performance depends on populating them evenly as well as on total capacity. Under-provision and workloads spill; over-provision every node and the capacity sits idle.

Disaggregated memory moves part of that decision from the server to the rack. See memory at rack scale and TORmem's memory-centric architecture.

How TORmem approaches L10

TORmem builds L10 servers to order on platforms from Tier-1 ODMs and OEMs: configured to the workload, firmware-matched and burn-in tested. They're available as single servers or as part of an L11 rack. Browse products, or send a procurement brief.

Frequently asked questions

What is L10 server integration?

L10 is the integration level at which a server is complete: CPUs, memory, GPUs, storage and network adapters installed, firmware and BIOS set, an operating system or image loaded, and the system tested and burned in. It ships as a ready-to-rack server.

What's the difference between a barebone (L6) and an L10 server?

A barebone (L6) is the platform only: chassis, motherboard, power supplies, fans and backplane, with no CPU, memory, GPUs or drives. An L10 server adds all of those, plus firmware configuration, an OS or image, and system testing and burn-in.

What tests does an L10 server go through?

Typically: functional checks that every component is detected at the right speed, diagnostics for GPUs, memory, storage and network adapters, and a burn-in under sustained load, followed by a final verification. Tools and pass criteria vary by build.

How long should GPU server burn-in run?

There is no single standard. Third-party guidance for AI infrastructure ranges from about 72 to 168 hours. TORmem builds to order and agrees burn-in duration and pass criteria with each client before the build.

What firmware and BIOS work happens at L10?

BIOS, BMC, GPU, network adapter and drive firmware are set to agreed versions, and BIOS options such as power profiles, NUMA and PCIe settings are configured for the workload. Matching versions across servers avoids hard-to-trace differences once they're racked together.

Can I buy L10 servers without rack integration?

Yes. L10 servers ship individually, ready to rack into existing infrastructure. Rack integration (L11) is optional and makes most sense for new racks, liquid-cooled designs, or teams without in-house integration capacity.

Does "L10" mean the same thing at every vendor?

No. The integration scale is an industry convention, not a standard. Some vendors use L10 for a rack shell, and in NVIDIA's supply chain it can mean a finished compute tray. Ask for an itemised list of what's installed and tested.

How should memory be configured in an L10 server?

Populate memory channels evenly (AMD EPYC 9005 processors have 12 DDR5 channels per socket) and size capacity to the workload. For memory-heavy workloads, disaggregated memory can add shared capacity at rack level instead of over-provisioning every server.