If PyTorch with DirectML on the NPU has stopped working on your Windows 11 AI laptop, you are not alone, and the problem is usually fixable without replacing hardware or reinstalling Windows. Running PyTorch through DirectML should let your NPU accelerate training and inference, so a device that never appears, errors on model load or runs slower than the CPU needs a careful driver and runtime check. The good news is that most failures come from a short list of causes — drivers, app versions, settings and requirements — and each one has a straightforward check.
This feature is one of the reasons many people bought an AI laptop in the first place, because it promises to do in seconds work that used to take minutes of manual effort. When it fails, it is tempting to assume the laptop is faulty or the feature was marketing hype, but in practice the underlying hardware is rarely the problem. More often, a single update, driver or setting has drifted out of step with the rest of the system, and restoring that alignment brings the feature back.
This guide walks through the likely causes first, then works through fixes in order, from the quickest checks to a clean reinstall. Along the way you will also find related fixes, such as dell xps 14 ai issues and copilot taskbar icon missing , which cover neighbouring AI features that share the same drivers and settings. Follow the steps in order, testing after each one, so you know exactly which fix solved it.
Why this happens
Power mode limiting the NPU: Battery saver and balanced power plans can throttle or park the NPU, so workloads run slowly or fall back to other hardware.
Missing runtime or framework: Frameworks such as DirectML or ONNX Runtime need matching runtimes, and a missing or mismatched runtime produces errors even with a healthy driver.
NPU disabled or hidden in firmware: A BIOS or firmware setting, or a firmware update gone wrong, can hide the NPU from Windows entirely.
Conflicting driver or background load: A second AI tool, a stuck inference job or a conflicting driver version can monopolise the NPU or report misleading usage figures.
Outdated or corrupted NPU driver: The NPU driver is what lets Windows and apps see and use the neural processor, so a stale, partially installed or corrupted driver breaks PyTorch with DirectML on the NPU first.
App not built to use the NPU: Not every AI app or model targets the NPU; some default to CPU or GPU, which looks like an NPU failure when it is actually the app’s design.
How to fix it
Work through these in order — diagnose first, then act.
Fix 1: Repair the app or runtime
When the driver is healthy but the app still fails, the app’s own runtime or installation is the next suspect, and repair is quicker than a full reinstall.
Open Settings, then Apps, and find the affected app.
Use Repair in its advanced options where available.
For framework tools, reinstall the matching runtime version.
Restart the laptop after the repair.
Run a small test workload to confirm the fix.
Fix 2: Clean-reinstall the NPU driver
A driver that was updated over a broken install can stay broken, so removing it fully and installing fresh resolves problems that simple updates do not. Microsoft’s own guidance at Microsoft Windows support is a good cross-check for the steps below.
Download the correct latest driver first, before removing anything.
In Device Manager, uninstall the NPU device, deleting driver software if offered.
Restart the laptop and let Windows redetect the hardware.
Install the downloaded driver and restart again.
Verify the NPU in Task Manager and retry PyTorch with DirectML on the NPU.
Fix 3: Check firmware and BIOS settings
If the NPU is missing from Windows entirely, firmware is involved, and a BIOS update or a toggled setting may be the only fix.
Check the laptop maker’s support page for a BIOS or firmware update.
Apply it only while plugged in, and do not interrupt it.
Enter BIOS setup and confirm any NPU or AI accelerator setting is enabled.
Save, exit and boot into Windows.
Check Device Manager and Task Manager for the NPU.
Fix 4: Verify with Task Manager and a test workload
Finally, confirm the fix with evidence rather than a feeling: a healthy NPU should appear in Task Manager and take load during an AI task.
Open Task Manager and select the NPU under Performance.
Run a small AI workload or benchmark.
Watch for NPU activity rising during the task.
If usage stays at zero, the app may not target the NPU — check its settings.
Record the result so you can spot a regression later.
Fix 5: Update the NPU driver from the vendor
Start with the driver itself, using the chip vendor or laptop maker rather than a generic updater, since NPU drivers are model-specific. For driver downloads, stick to the official vendor or laptop-maker pages linked in this guide rather than third-party driver sites.
Open Device Manager and expand Neural processors.
Note the exact NPU name and current driver version.
Download the latest driver from the vendor or laptop maker’s support page.
Install it and restart the laptop.
Confirm the NPU appears in Task Manager without a warning icon.
Fix 6: Set a high-performance power plan
Power saving is a quiet NPU killer, so ruling it out early prevents chasing driver problems that are really power problems.
Open Settings, then System, then Power.
Set Power mode to Best performance while plugged in.
Plug in the charger for testing, even if you normally run on battery.
Close other AI apps that might be using the NPU.
Retry PyTorch with DirectML on the NPU and compare speed or behaviour.
Closing
Once PyTorch with DirectML on the NPU is working again, keep it that way by installing app and driver updates promptly, staying on a current Windows build and checking Task Manager if performance ever slips. If a future update breaks it again, you now know which checks to repeat first. For a related older fix, see recall snapshot retention settings , and if the problem persists across every AI feature, treat that as a sign to check drivers and firmware as a whole rather than one app at a time.
