What is L11 rack integration?

Where L11 sits on the integration scale

Server manufacturing is usually described on an L1–L12 scale, where each level includes everything below it. AMAX publishes the most widely cited version. For buying AI infrastructure, four levels matter:

What's inside an L11 rack

An L11 rack contains everything it needs to run as a unit once it is connected to site power, cooling and the upstream network. Typical scope:

• Servers built and tested at L10: GPU, CPU, storage or memory nodes, mounted in their planned positions.
• Top-of-rack switching for the compute fabric (InfiniBand or RoCE Ethernet), plus a separate management network for the BMCs.
• Optics and cabling: DACs, AOCs or transceivers and fibre, routed, labelled and documented port to port.
• Power distribution: PDUs sized to the rack's power budget, with inputs matched to the site's feeds.
• Liquid-cooling distribution for liquid-cooled designs: rack manifolds, hoses and quick-disconnects to the cold plates, an in-rack CDU or a connection to a facility CDU, and leak detection.
• Aligned configuration: firmware, BIOS and BMC settings matched across every node in the rack.

Where cooling work falls varies. Gigabyte places cold plates and quick-disconnects at L10, and manifolds, CDUs and leak detection at L11.

How an L11 rack is tested

L11 testing exists to catch problems no single-server test can see: a cable in the wrong port, a link negotiating at the wrong speed, a PDU that trips when every GPU hits peak at once, a hot spot that only appears with the rack fully loaded.

Published rack-scale validation frameworks, such as ASUS's for NVIDIA NVL72 racks, run in a similar order:

1. Node checks: each server confirmed healthy in its final position, carried over from L10.
2. Power-on and power cycling: every node, switch and PDU, including behaviour under load.
3. Firmware and BIOS alignment: versions and settings confirmed identical across nodes.
4. BMC and management network: every controller reachable over Redfish or IPMI.
5. Cabling validation: every link up at the expected speed and matched against the cabling plan.
6. Fabric and collective tests: NCCL tests across GPUs, HPL or similar for sustained load, and vendor diagnostics such as NVIDIA DCGM.
7. Thermal and load soak: the full rack held at load to confirm that power and cooling hold.

Duration and pass criteria vary by build. Third-party guidance suggests multi-day burn-in with results within 5–10% of target, and Oracle's NVIDIA Exemplar Cloud bar is within 5% of NVIDIA's reference workloads. TORmem builds to order, so the test plan and pass criteria are agreed with each client before the build starts.

Reading collective-test results is a skill of its own. See how to read NCCL error logs and why a GPU fabric can run at half bandwidth.

What you should receive with the rack

The documents that ship with an L11 rack decide how fast it can be brought up on site. NVIDIA's rack bring-up checklist for Mission Control expects a rack inventory file "obtained from factory after L11 testing" and point-to-point cabling documentation with MAC addresses before provisioning starts.

• Rack inventory: every node, switch and major component by serial number and position.
• Cabling map: port-to-port connections, with MAC addresses for managed interfaces.
• Firmware and BIOS baseline: the versions and settings applied to every node.
• Test reports: what was run, for how long, and the results against the agreed criteria.
• As-built documentation: rack elevation, power plan and PDU assignments.

TORmem specifies the handover package with each client. Whoever you buy from, get the list in writing as part of the quote. It's much harder to ask for after the rack has shipped.

Power and cooling at rack scale

AI racks sit far outside traditional data center norms. Uptime Institute's 2026 survey puts the average rack at about 11 kW. An NVIDIA GB200 NVL72 rack draws around 120 kW, and HPE lists 132 kW nominal and recommends provisioning for a peak around 192 kW. GB300 racks reach up to 142 kW.

Air cooling tops out at roughly 30–50 kW per rack, which is vendor consensus rather than a standard. Most current AI racks are therefore liquid-cooled.

• Provision power for the rack's peak draw, not its nominal figure.
• Confirm the site's voltage and feed configuration before the PDUs are specified.
• For liquid-cooled racks, agree whether the CDU sits in the rack or in the facility, and pressure- and leak-test before shipping.
• If the site doesn't have the power yet, that's a project in its own right. TORmem coordinates on-site generation through energy partners. See procurement.

Shipping and receiving a loaded rack

A fully integrated AI rack is heavy. NVIDIA's DGX GB200 NVL72 weighs roughly 1.36 tonnes (3,000 lb). Racks and shock pallets carry two different ratings, static and ship-loaded, and they are not the same: one example pallet is rated for 1,250 lb in transit against 3,000 lb static. A rack that stands safely on the floor may not be rated to travel fully loaded.

• Check the ship-loaded rating of the rack and pallet against the loaded weight.
• Use shock-absorbing crates and air-ride freight. Supermicro's integration service lists both.
• Walk the receiving path: dock height, door and lift dimensions, corridor turns and floor loading all the way to the hall.
• Decide who uncrates, positions and connects the rack, and how it is re-tested on arrival.

There is no published industry standard for AI-rack floor loading. Get the figures from the facility before the rack ships.

Timeline and cost drivers

There is no standard turnaround for L11. Published figures run from "days rather than weeks" for deployment to two to eight weeks for integration. In practice the schedule is set by the longest-lead component: one reported build waited 14 weeks for a single part.

What drives cost and time:

• Component availability, especially GPUs, switches and optics
• Test scope and duration
• Cabling and optics count, which grows with the fabric design
• Liquid-cooling parts and leak testing
• Crating, freight and insurance
• On-site installation and acceptance testing

A quote that prices each of these separately is far easier to compare than a single rack price.

Memory at rack scale

Most L11 discussions stop at GPUs, network and cooling. Memory is usually fixed server by server at L10, and whatever a server can't hold, its workload can't use. Disaggregated memory moves part of that capacity into a shared pool in the rack, reached over the fabric, so memory is sized for the rack instead of over-provisioned in every node.

That makes memory an L11 decision. A memory appliance needs rack space, power and fabric ports in the rack plan, and its performance has to be validated with the rest of the rack. See TORmem's memory-centric architecture, CXL vs RDMA over Ethernet and the M1000 memory appliance.

How TORmem approaches L11

TORmem builds to order. Rack layout, components, network fabric, cooling approach, test plan, handover documents and delivery are specified with each client, then built and validated before the rack ships. Alongside GPU compute, networking and storage, TORmem can integrate its own disaggregated memory systems into the rack. See L11 rack integration for scope, or send a procurement brief.

Frequently asked questions

What is L11 rack integration?

L11 is the integration level at which configured servers are racked, cabled to top-of-rack switches and power distribution, and tested together as one system before shipping. The customer receives a working, validated rack instead of individual servers to assemble on site.

What's included in an L11 rack?

Typically: L10 servers mounted in the rack, top-of-rack switches for the compute fabric and management network, optics and cabling, PDUs, and firmware and BIOS settings aligned across nodes. Liquid-cooled designs add rack manifolds, a CDU or CDU connection, and leak detection. Scope varies by vendor, so ask for an itemised list.

How is an L11 rack tested before it ships?

In stages: node health checks, rack power-on and power cycling, firmware and BIOS alignment, BMC reachability, cabling validation against the plan, fabric and collective tests such as NCCL tests, and a thermal and load soak. Duration and pass criteria are agreed for each build.

What documentation should come with an L11 rack?

At minimum: a rack inventory by serial number and position, a port-to-port cabling map with MAC addresses, the firmware and BIOS baseline, and test reports. NVIDIA's Mission Control bring-up process expects an inventory file from L11 testing and point-to-point cabling documentation.

How long does L11 integration take?

There is no standard figure. Published integrator ranges run from days to around eight weeks, and the schedule is usually set by the longest-lead component, such as GPUs, switches or optics, rather than by assembly time.

How much power does an AI rack draw, and when is liquid cooling required?

NVIDIA GB200 NVL72 racks draw around 120–132 kW, with higher peaks, against an industry average near 11 kW per rack. Air cooling generally tops out around 30–50 kW per rack, so racks above that range are typically liquid-cooled.

How are fully integrated racks shipped?

In shock-absorbing crates on air-ride freight. Check that the rack and pallet are rated for their loaded weight in transit, not only static, and confirm the receiving path — dock, doors, lifts and floor loading — before the rack leaves the factory.

What's the difference between L11 and L12?

L11 is a single rack, integrated and tested. L12 connects multiple racks into a cluster, adding cross-rack cabling and cluster-level testing, and sometimes cluster management software.

Can disaggregated memory be integrated at L11?

Yes. A memory appliance takes rack space, power and fabric ports like any other node, so it belongs in the L11 rack plan and is validated with the rest of the rack. TORmem designs disaggregated memory systems and integrates them into L11 builds.