How I Got an NVIDIA L40 Working as a Display GPU in Windows 11

The complete desktop workstation conversion, including the mistakes that made it harder than it needed to be.

I could find plenty of information about using an NVIDIA L40 as a compute accelerator. Practical information about making its DisplayPort outputs drive a normal Windows 11 desktop was much harder to find. After a long sequence of firmware changes, driver experiments, black screens, cooling work, and one embarrassingly important hardware discovery, I got it working.

This is not my gaming machine. My gaming system uses an RTX 5090. The L40 went into a separate desktop workstation that previously used an RTX 3090 Founders Edition. The point of this build was not gaming performance. I wanted the L40's 48 GB of ECC memory to test local diffusion models in ComfyUI, computer vision inference, local large language models such as Hermes, and other workloads that benefit from a large pool of VRAM.

The L40 used in this project was made available to me temporarily for evaluation. I did not purchase it, and this article is not sponsored. The testing, observations, and opinions are my own.

What follows is the complete version of the project. I have included the successful path, the dead ends, the cooling configuration, the commands I used, and the validation results. Some of the failures turned out to be more useful than a clean installation would have been because they exposed the difference between the L40's firmware display mode, its Windows driver model, and the physical display connection.

Last updated: August 28, 2026
Final result: NVIDIA L40 driving a Dell U3818DW at 3840 × 1600 through DisplayPort, in WDDM mode, with temperature-controlled external fans.

The completed NVIDIA L40 desktop workstation beside the Dell U3818DW

The completed workstation with the Dell U3818DW connected directly to the L40.

Important: This is a case study, not an NVIDIA-supported reference build. The L40 is a passive data-center GPU. It must have forced airflow through its heatsink, and changing its board display mode carries firmware risk. Back up important data, create a restore point, keep a fallback display path available, and proceed at your own risk.

Short version: what finally worked

The successful sequence was:

Build and test a forced-air shroud for the passive L40.
Shut down, remove AC power, remove the RTX 3090 FE, and install the L40.
Fully seat the vertical PCIe riser at both the motherboard and GPU ends. A loose motherboard-side riser connection was the cause of the most serious instability encountered in this build.
Use a correctly wired, PSU-compatible PCIe power solution for the L40's 16-pin connector; never mix modular PSU cables from different models.
Keep the monitor connected to the Ryzen 9 7950X's motherboard video output as a recovery path during setup.
In UEFI, use UEFI boot with CSM disabled, Above 4G Decoding enabled, and Re-Size BAR enabled or Auto. Leave the primary-display preference on iGPU or Auto until everything is stable.
Set the L40 board to physical_display_enabled_256MB_bar1 with NVIDIA Display Mode Selector, then reboot. Do not select the 8 GB BAR1 display mode merely because it sounds larger.
Install NVIDIA RTX Enterprise DCH driver 596.86.
If nvidia-smi reports MCDM, run this in an elevated PowerShell window:
```
nvidia-smi -i 0 -dm 0
```
Reboot. Confirm that nvidia-smi now reports WDDM.
Confirm that NVIDIA Control Panel exposes Manage 3D settings and lists the four DisplayPort connectors under View system topology.
Shut down, move the DisplayPort cable from the motherboard to the L40, and power on.
Confirm that nvidia-smi reports WDDM and Disp.A On.
Configure Fan Control to drive the two shroud fans from the L40 temperature reported by HWiNFO.
Validate the system under a real GPU load while watching temperature, power, fan speed, PCIe errors, and stability.

That final configuration worked. The L40 displayed Windows directly, idled at about 40 W after settling, and completed Blender Benchmark at approximately 290 W without exceeding 69 °C.

The hardware used

Component	Configuration in this build
GPU being removed	NVIDIA GeForce RTX 3090 Founders Edition
GPU being installed	NVIDIA L40, 48 GB ECC GDDR6, passive, 300 W
CPU	AMD Ryzen 9 7950X
Motherboard	ASRock X670E Steel Legend, current UEFI/BIOS
Power supply	EVGA SuperNOVA 1000 GT
Case	Lian Li O11 Dynamic EVO XL
GPU mounting	Lian Li vertical mount and PCIe riser
L40 cooling	Shroud model from Printables.com, 3D printed in ASA Glass Filled (ASA-GF), with two ARCTIC S4028-15K 40 × 28 mm PWM fans
Monitor	Dell U3818DW, 3840 × 1600 at 60 Hz, DisplayPort 1.2
Operating system	Windows 11
Final NVIDIA driver	RTX Enterprise DCH 596.86
Monitoring/control	HWiNFO64 8.50, Fan Control V273, FanControl.HWInfo plug-in

NVIDIA L40 installed vertically in the workstation

The passive L40 installed in the Lian Li vertical mount. The shroud and its two high static pressure fans sit at the card's intake end.

The official NVIDIA L40 page specifies a passive, dual-slot, 300 W card with 48 GB ECC GDDR6, one PCIe CEM5 16-pin power connector, and four DisplayPort 1.4a outputs. “Passive” does not mean that the card can run without cooling; it means the host chassis is expected to provide the airflow.

The two ARCTIC S4028-15K fans are 4-pin PWM fans rated for 1,400–15,000 RPM, 26.5 mmH₂O static pressure, and 0.47 A at 12 V each. Each fan was connected to its own motherboard chassis-fan header. ARCTIC advises using a header capable of at least 0.5 A for one fan.

Before starting

1. Preserve a recovery display path

The Ryzen 9 7950X used here has integrated AMD Radeon Graphics. During the risky parts of the process, the monitor was connected to a motherboard video output. This made it possible to repair drivers and settings even when the L40 produced no Windows image.

Do not disable the iGPU during initial setup. Leave the BIOS primary-display option on iGPU or Auto until the L40 is confirmed working. Even after success, keeping the iGPU enabled provides a valuable recovery path.

2. Create a Windows restore point

Open Start and search for Create a restore point.
Open System Protection.
Select the Windows system drive and choose Configure if protection is off.
Enable system protection and allocate reasonable disk space.
Click Create, name the point, and wait for confirmation.

A restore point is not a substitute for a file backup, but it can help reverse driver and registry changes.

3. Download everything before removing the old GPU

Have these available locally:

The intended NVIDIA driver package.
NVIDIA Display Mode Selector from the official NVIDIA developer page.
HWiNFO64.
Fan Control from its official GitHub releases page.
The FanControl.HWInfo plug-in.
Optionally, Display Driver Uninstaller (DDU) for recovery, but not as an automatic first step.

4. Record a baseline

Before disassembly, record:

BIOS version and relevant PCIe settings.
Existing NVIDIA driver version.
A photo of all cabling and the vertical-riser installation.
Idle and load temperatures for the old GPU.
Which motherboard headers are free.

5. Understand the three independent layers

Much of the confusion in this project came from treating three separate settings as if they were one:

L40 board firmware display mode
        ↓
Windows NVIDIA driver model: MCDM or WDDM
        ↓
Physical DisplayPort link and active Windows display

Board display mode is stored in the card's EEPROM and is changed by Display Mode Selector.
Driver model is controlled by the installed Windows driver and nvidia-smi. MCDM is a compute-oriented model; WDDM is required for normal Windows graphics/display use.
DisplayPort link is the actual cable, monitor handshake, EDID, resolution, and active desktop connection.

A card can have physical display mode enabled in firmware while still loading as MCDM and producing no Windows desktop.

Phase 1: build and verify the cooling solution

The L40 has no onboard blower. Do not install it and begin stress testing before the shroud and forced airflow are ready.

Shroud goals

Seal the fans to the heatsink inlet so air is pushed through the fin stack rather than escaping around it.
Avoid recirculating hot exhaust back into the inlet.
Keep wiring clear of fan blades.
Use secure fasteners and a temperature-tolerant print material.
Confirm airflow direction before fitting the shroud.

The shroud model used for this build came from Printables.com and was 3D printed in ASA Glass Filled (ASA-GF). Two ARCTIC S4028-15K fans were mounted to the shroud. Both fan headers were set for PWM control. Each fan used its own chassis-fan header, avoiding current-sharing questions and allowing independent RPM feedback.

Safe BIOS fallback

Before Fan Control loads, the motherboard owns the fans. Set the two headers to PWM mode and choose a nonzero, safe fallback rather than a stop mode. A fixed moderate duty such as roughly 40–50% is reasonable during setup if the noise is tolerable. Once Windows and Fan Control are known to start reliably, the BIOS fallback remains insurance if the application or sensor plug-in fails.

Phase 2: remove the RTX 3090 FE and install the L40

Shut down Windows completely.
Turn off the PSU and disconnect the AC cord.
Press the case power button briefly to discharge residual power.
Remove the side panels needed for safe access.
Photograph the existing GPU, power cables, riser, and fan routing.
Disconnect the RTX 3090 FE power connector without pulling on the wires.
Support the card, remove its bracket/mount fasteners, release the PCIe latch if directly slotted, and remove it.
Inspect the vertical riser connector and cable for damage or strain.
Install the L40 in the vertical mount and secure it mechanically.
Reseat and verify both ends of the PCIe riser. This is crucial. The motherboard-side connection in this build was initially loose, which later caused Code 43, severe Windows sluggishness, and misleading driver symptoms.
Connect the correct L40 power cable or adapter. Fully seat the 16-pin plug; no contacts should be visible at the mating edge. Avoid a hard bend immediately at the connector.
Do not use a modular cable simply because it physically fits the PSU. Modular pinouts are not universal. Use only a cable explicitly wired for the EVGA PSU family in question, or the correct adapter fed from appropriate PCIe outputs.
Connect both shroud fans to the selected chassis-fan headers.
Verify that the shroud cannot contact the fan blades and that its outlet is unobstructed.
For the first boot, keep the display cable connected to the motherboard output.

Power capacity

The single-L40 configuration was appropriate for the 1000 W EVGA unit used here. NVIDIA rates the L40 at 300 W, and the card reached about 290 W during Blender Benchmark. PSU wattage alone is not enough: cable compatibility, connector seating, rail/cable loading, and transient behavior also matter.

Installing a second L40 was not pursued. It would materially increase power and cooling requirements, and the X670E Steel Legend's second full-length slot is electrically much more limited than its primary slot. Two cards also add software and airflow complexity that was not justified for this workstation.

Phase 3: UEFI/BIOS checks

The working configuration used:

CSM: Disabled
Boot mode: UEFI
Above 4G Decoding: Enabled
Re-Size BAR: Enabled or Auto
Primary graphics: iGPU or Auto during setup
Integrated graphics: Enabled/available as a fallback

Do not invent a large manual MMIO aperture or force a “40-bit” value unless the motherboard documentation and a specific platform requirement call for it. Above 4G Decoding is the relevant setting; arbitrary MMIO changes can create new problems.

These BIOS settings do not need to be repeated after every NVIDIA driver installation. The display-mode firmware setting and the Windows driver model are separate from Re-Size BAR.

Phase 4: first Windows boot and hardware verification

With the monitor still on the motherboard output, boot Windows and open an elevated PowerShell window.

Run:

nvidia-smi -L

Expected result: an NVIDIA L40 is listed. Do not publish the UUID shown by this command.

Then run:

nvidia-smi

Check:

GPU name is NVIDIA L40.
Temperature is plausible.
Power cap is 300 W.
Approximately 48 GB of memory is shown.
No Device Manager warning icon or Code 43 is present.

At the beginning of this build, the Data Center 610.88 driver detected the L40 correctly but loaded it as MCDM, with Disp.A Off. Detection alone did not mean the DisplayPort outputs were ready for the Windows desktop.

What `Disp.A` means

Disp.A Off means no display is currently active on that GPU.
It does not, by itself, prove that the card's ports are permanently disabled.
Disp.A On after connecting the display to the L40 is the desired final result.

Phase 5: enable the L40's physical display mode

NVIDIA describes Display Mode Selector as a firmware tool for switching supported boards such as the L40 between display-enabled and display-disabled modes. NVIDIA also warns that an unsuitable mode or interrupted update can make a board unusable. Read the warning on the official Display Mode Selector page before proceeding.

Precautions

Use stable wall power and do not interrupt the operation.
Close GPU workloads and monitoring utilities.
Use the official tool and run it from an elevated terminal.
If Windows Memory Integrity produces a warning, do not assume the operation failed; read the tool's exact output.
Do not flash a VBIOS from another card.
Record the original mode before changing anything.

From the directory containing the utility, run:

./displaymodeselector.exe --gpumode

The tool presented these choices:

physical_display_enabled_256MB_bar1
physical_display_disabled
physical_display_enabled_8GB_bar1

For this build, the selected target was:

physical_display_enabled_256MB_bar1

The tool reported a successful EEPROM update and required a reboot. The interactive menu labeled the selected item as option 0, while the completion text referred to the applied firmware setting as “Mode 1.” Preserve the descriptive mode name in your notes; it is less ambiguous than the menu number.

Why not 8 GB BAR1?

The 8 GB option is not a generic performance upgrade. It changes the board's display-mode resource layout and may require host/platform support. The 256 MB BAR1 mode was the conservative, working choice here. Re-Size BAR in BIOS is a separate setting.

Phase 6: choose the Windows driver and force WDDM

The unsuccessful path

With the Data Center 610.88 driver, the card could compute and render, and it could show pre-Windows firmware screens. After display mode was enabled and the driver model was changed, however, the Dell monitor still went black when Windows loaded. Trying different ports, Windows Detect, DisplayPort 1.1 mode, and an EDID override did not fix the actual Windows link.

An early attempt to install an RTX driver resulted in Code 43 and an almost unusably slow desktop. That looked like a driver incompatibility, but the later physical inspection revealed a loose PCIe riser connection at the motherboard. Once the riser was fully seated, the same general RTX Enterprise path worked. The lesson is simple: verify the physical PCIe link and power before cycling through drivers.

The successful driver

The final working package was NVIDIA RTX Enterprise DCH 596.86. NVIDIA's RTX Enterprise branch history identifies 596.86 in the R595 branch.

The installer was allowed to install the graphics driver. The HD Audio component is only necessary if audio will travel through DisplayPort/HDMI; it does not enable the video signal itself.

After installation, Device Manager showed the L40 without a yellow warning, but nvidia-smi initially showed:

Driver-Model: MCDM
Disp.A: Off

NVIDIA Control Panel at that point was compute-oriented: Manage 3D settings was missing, and the card was described as dedicated for compute needs. That behavior was consistent with MCDM, not proof of a failed installation.

Switch MCDM to WDDM

Open PowerShell as Administrator and run:

nvidia-smi -i 0 -dm 0

The command returned that the driver model had been set to WDDM and that a reboot was required.

Reboot before judging the result.

After reboot, run:

nvidia-smi

The desired result is:

Driver-Model: WDDM

NVIDIA's NVML documentation defines WDDM as the Windows display driver model and MCDM as the Microsoft compute driver model. In practical terms for this build, WDDM was necessary for a normal Windows desktop and for the full NVIDIA Control Panel graphics pages.

Confirm the control-panel change

After the WDDM reboot, NVIDIA Control Panel exposed:

Adjust image settings with preview
Manage 3D settings
Video controls
Set PhysX configuration
View system topology

Under Workstation → View system topology, the L40 showed four DisplayPort entries. With the cable still on the motherboard, each correctly showed Not connected. This verified that the driver was enumerating the display engines and connectors.

Support caveat: NVIDIA's public Data Center documentation does not present a consumer-desktop L40 as a normal supported GeForce-style configuration, and supported WDDM/vGPU scenarios may have different licensing and platform requirements. Treat this as a documented working configuration, not a guarantee that every L40, firmware revision, driver branch, motherboard, or monitor will behave identically.

Phase 7: move the monitor to the L40

Although DisplayPort is designed for hot-plugging, a full shutdown was used to minimize variables.

Confirm nvidia-smi reports WDDM.
Confirm Device Manager has no warning on the L40.
Confirm NVIDIA Control Panel lists all four DP connectors.
Shut down Windows.
Turn the monitor off.
Move the DisplayPort cable from the motherboard to one L40 port.
Turn on the monitor and select the correct DisplayPort input.
Power on the PC.

The system displayed the UEFI/boot screens and then the Windows desktop from the L40.

Run:

nvidia-smi

The final confirmation showed:

Driver-Model: WDDM
Disp.A: On

nvidia-smi showing the NVIDIA L40 in WDDM mode with an active display

The result I had been trying to reach: WDDM, Disp.A On, the full 48 GB memory pool, and Windows running from the L40.

Windows desktop processes appeared under the L40, and approximately 1.9 GB of VRAM was in use. Immediately after switching, the card was temporarily around 80 W in P0; after the desktop settled, it dropped to about 40 W in P8 at approximately 43 °C. That transient was normal for this system.

BIOS preference after success

The L40 worked while the BIOS primary-display preference remained on iGPU/Auto. There was no need to change it to External Graphics merely to obtain Windows output. Leaving the iGPU available is useful for recovery. A user may later test External/PCIe as primary, but it is optional and should be changed only after the working state is documented.

Phase 8: configure quiet, temperature-based L40 cooling

The L40's onboard fan reading is N/A because the card is passive. The shroud fans are motherboard fans, so Fan Control must control motherboard headers using an L40 temperature sensor.

Why HWiNFO was needed

With the Data Center/enterprise configuration, Fan Control's native NVAPI source repeatedly produced:

NVApi could not initialize or has no sensors.

HWiNFO could read the L40 temperature correctly. The FanControl.HWInfo plug-in then exposed that temperature to Fan Control.

HWiNFO setup

Launch HWiNFO in Sensors-only mode.
Open the L40 sensor group and confirm GPU Temperature is present.
Open sensor settings for that entry.
Enable Report to Gadget / reporting for the L40 GPU temperature.
Configure HWiNFO to start with Windows, minimized, with sensors active.
The free, non-commercial HWiNFO edition was sufficient for this personal setup.

The HWiNFO registry sensor export was verified separately before troubleshooting Fan Control.

Fan Control source settings

In Fan Control:

Enable the HWiNFO plug-in.
Disable the NVAPI sensor source if it repeatedly fails on the L40.
Leave NVIDIA 0% fan curve override disabled; it applies to supported onboard NVIDIA fans, not these motherboard-connected shroud fans.
Use a delayed startup. Thirty seconds worked and gave HWiNFO time to expose its sensor.
Start minimized at user logon.
Retain a safe BIOS fan fallback in case Windows control fails.

Identify the correct headers

Change one detected control at a time while watching and listening to the fans. In this build:

System Fan #1 became L40 Fan #1.
System Fan #4 became L40 Fan #2.

At a 20% PWM command, the fans reported approximately 3,690 and 3,670 RPM after reboot.

Do not assume PWM percentage maps linearly to rated RPM. A 50% duty reading near 10,000 RPM on a nominal 15,000 RPM fan is plausible. The controller, fan electronics, static pressure, tachometer scaling, and operating point all affect the result. Full assisted calibration was not necessary once the two headers and safe minimum were known.

Working L40 fan curve

Create a Graph curve named L40 Cooling, with the HWiNFO L40 GPU Temperature as its source:

L40 temperature	Fan command
30 °C	20%
40 °C	22%
50 °C	30%
60 °C	42%
68 °C	58%
74 °C	75%
80 °C	100%

Additional settings:

Hysteresis: 3 °C
Response time: 3 seconds
Symmetric: enabled
Ignore hysteresis at minimum and maximum temperatures: enabled

Assign L40 Cooling to both L40 fan controls. Verify visually that both control cards show the curve name, the same commanded percentage, and nonzero RPM.

Fan Control using the L40 temperature to manage both shroud fans

Both shroud fans following the same L40 temperature curve through the HWiNFO sensor source.

This curve is a starting point, not a universal prescription. Shroud sealing, room temperature, case airflow, workload, and noise tolerance differ. The important rule is to validate under sustained load and ensure that neither fan can silently stop.

Phase 9: validation

Live monitoring command

In PowerShell:

nvidia-smi --query-gpu=temperature.gpu,power.draw --format=csv -l 2

This prints temperature and power every two seconds. Stop it with Ctrl+C.

Also monitor in HWiNFO:

L40 GPU temperature
GPU power
GPU clocks and load
Available/allocated GPU memory
PCI Express error counters
PCIe link speed
Both shroud-fan RPM readings

Blender Benchmark result

The official Blender Benchmark 5.2.0 was run with OptiX. Results were:

Scene	Samples per minute
Monster	5,139.096054
Junkshop	3,472.023028
Classroom	2,871.048188
Total score	11,482.17

During the run:

Peak GPU power was roughly 290 W.
Maximum observed GPU temperature was 69 °C.
Both shroud fans visibly ramped up and down with the L40 curve.
The benchmark completed successfully.

HWiNFO validation data for the NVIDIA L40

HWiNFO captured a 79.7 W peak in this particular sensor window. The longer Blender validation run reached roughly 290 W and topped out at 69 °C.

Blender Open Data described the result as “top 0%.” That is favorable. It means the rank was in a fraction of the top 1%, rounded to a whole percentage, not that the score was zero or invalid. Cross-generation comparisons should still use the same Blender version, backend, scenes, and settings.

Reboot validation

After a final reboot:

The L40 continued to drive the monitor.
HWiNFO started minimized.
Fan Control started without the earlier initialization error.
The HWiNFO L40 temperature source was available.
Both L40 fans followed the curve at about 20% and approximately 3,700 RPM.
Idle temperature was around 32–33 °C when the display was on the motherboard, and around the low 40s after the L40 became the active display, depending on desktop activity.

Troubleshooting guide

Symptom: `nvidia-smi` shows MCDM

Meaning: The L40 is loaded as a Microsoft compute device. Normal Windows graphics pages and display output may be unavailable.

Action: In an elevated PowerShell window:

nvidia-smi -i 0 -dm 0

Then reboot and verify that the result is WDDM.

Symptom: UEFI/BIOS appears through the L40, but Windows is black

This proves that the physical output can produce a pre-OS signal, but not that the Windows driver is in the correct mode.

Check, in order:

Motherboard fallback display works.
Device Manager has no Code 43.
The riser is fully seated at both ends.
GPU power is fully seated and correctly wired.
The L40 board is in physical-display-enabled 256 MB BAR1 mode.
nvidia-smi reports WDDM after reboot.
NVIDIA Control Panel exposes Manage 3D settings and four DP ports.
Only then retry the L40 DisplayPort connection.

Symptom: Device Manager Code 43 and severe Windows lag

Do not immediately assume the driver package is wrong. Power down and inspect:

Motherboard end of the PCIe riser
GPU end of the riser
16-pin GPU power connector
PSU-side modular connectors
Cable damage or extreme bending

In this build, reseating the loose motherboard-side riser resolved the hardware fault that had masqueraded as a driver problem.

Symptom: NVIDIA Control Panel lacks Manage 3D settings

Check nvidia-smi. If the L40 is in MCDM, the reduced compute/workstation interface is expected. Switch to WDDM, reboot, and check again.

Symptom: Control Panel lists all four ports as “Not connected”

That is correct while the cable is connected to the motherboard. After connecting the DP cable to the L40 and obtaining an active Windows image, Disp.A should change to On.

Symptom: Windows thinks there are two identical displays

This occurred during the unsuccessful motherboard-output plus EDID-override experiments. In Settings → System → Display, select the unwanted display and choose Disconnect this display or remove it from the desktop. Avoid forcing an EDID override unless basic physical, driver-model, and connector checks have already passed. Remove experimental EDID overrides when returning to a clean configuration.

Symptom: Fan Control says NVAPI cannot initialize

Disable the failing NVAPI source, confirm HWiNFO can see the L40 temperature, enable reporting for that sensor, load the FanControl.HWInfo plug-in, and use a delayed startup. The error does not mean the motherboard fan outputs are unusable.

Symptom: fans are loud at 20–50%

These are server-class 15,000 RPM fans. Even low PWM duty can be audible. Confirm the shroud is sealed, both RPM readings are valid, and the curve does not jump unnecessarily. Do not lower the minimum below the fans' reliable starting/running point simply to chase silence.

What did not solve the Windows DisplayPort problem

The following were tried before the final successful sequence and did not fix the underlying issue:

Forcing the Dell monitor to DisplayPort 1.1.
Windows Detect display.
Trying different L40 DisplayPort sockets without first fixing the driver/hardware state.
Exporting and applying a monitor EDID override.
Treating Disp.A Off as proof that the ports were disabled.
Trying to solve MCDM behavior only through Windows display layout.
Repeatedly swapping drivers before rechecking the PCIe riser.
Considering physical_display_enabled_8GB_bar1 as a generic upgrade.

The resolution and DP 1.2 capability of the Dell U3818DW were not the root cause. Once the physical PCIe connection, RTX Enterprise driver, and WDDM state were correct, the same monitor and cable worked at 3840 × 1600.

Final known-good configuration

Hardware

One NVIDIA L40 in the primary x16 slot through the Lian Li vertical riser
Riser fully seated at both ends
Correct, fully seated 16-pin GPU power connection
Two 40 × 28 mm high-static-pressure PWM fans in a sealed shroud
One fan per motherboard chassis-fan header
Dell U3818DW connected directly to an L40 DisplayPort
EVGA SuperNOVA 1000 GT powering a single L40 system

Firmware and BIOS

L40 mode: physical_display_enabled_256MB_bar1
UEFI boot; CSM disabled
Above 4G Decoding enabled
Re-Size BAR enabled or Auto
BIOS primary graphics left on iGPU/Auto
Integrated graphics retained for emergency recovery

Windows and NVIDIA

NVIDIA RTX Enterprise DCH 596.86
L40 driver model: WDDM
Disp.A On with the DP cable attached
Full NVIDIA Control Panel graphics pages present
Device Manager clean, no Code 43

Cooling and monitoring

HWiNFO running sensors minimized at startup
HWiNFO L40 GPU temperature exported to Fan Control
Fan Control delayed startup by 30 seconds
Both shroud fans assigned to the L40 Cooling graph
Safe nonzero BIOS fallback retained

Lessons worth carrying into another build

Fix the physical layer first. A partly seated riser can look exactly like a driver failure.
Firmware display mode and Windows driver mode are separate. A display-enabled L40 can still load as MCDM.
Reboot after changing the driver model. The command is not fully applied until reboot.
Use the iGPU as a lifeline. It prevented a black-screen experiment from becoming a recovery crisis.
Do not stress a passive accelerator without forced airflow. A printed shroud is functional infrastructure, not decoration.
PWM percentage is not RPM percentage. Calibrate for safe temperature and acoustics, not mathematical neatness.
Change one variable at a time and record the state. Driver, firmware, BIOS, EDID, cable, and display layout changes can otherwise obscure the real cause.
Avoid unnecessary firmware experiments. The 256 MB BAR1 mode worked; the 8 GB option was not needed.
Preserve a rollback path. Keep installers, restore points, and the motherboard display available until several clean reboots and load tests have passed.