How to Fix Windows 11 Copilot+ NPU Not Detected (2026): BIOS Switches, NPU Drivers, ONNX Runtime, and AI App Compatibility

Copilot+ PCs are built around local AI acceleration—but in 2026 it’s still surprisingly common to see “NPU not detected”, an AI app that only offers CPU/GPU backends, or Windows features that refuse to light up even though you bought a Snapdragon X, Ryzen AI, or Core Ultra machine.

This guide is for power users: we’ll walk through the exact layers that commonly break—BIOS toggles, OEM NPU drivers, Windows components, and runtime stacks like ONNX Runtime / DirectML—and how to prove (not guess) where the failure is.

Key Takeaways (Read This First)

  • Don’t start with your AI app. First confirm the NPU exists at the firmware + Windows device layer (Device Manager + dxdiag + Event Viewer).
  • Most “NPU missing” cases are driver-stack issues: wrong OEM package, Windows Update overwrote a vendor driver, or a stale AI runtime.
  • BIOS/UEFI can hide or disable the NPU path via AI/Neural/UMA memory/security toggles—especially after a BIOS update/reset.
  • ONNX Runtime provider choice matters. Many apps silently fall back to CPU if the NPU execution provider isn’t present or the model isn’t NPU-friendly.
  • Compatibility isn’t universal. Some models (opset, quantization, dynamic shapes) won’t run on NPU even if detection is working.

1) Confirm Your PC Actually Meets Copilot+ / NPU Requirements (Fast Sanity Check)

Copilot+ PC branding typically indicates an NPU capable of ~40+ TOPS (varies by platform/year). However, identification inside Windows can still fail if you’re on the wrong OS build, in an enterprise image, or using a non-standard hypervisor setup.

  • Windows Build: Go to Settings → System → About. Ensure you’re on a current Windows 11 build with the latest cumulative updates.
  • Virtualization: If you’re running Windows inside a VM, you likely won’t see the NPU at all. NPUs are not broadly virtualized like GPUs.
  • Corporate/Managed devices: Security baselines can restrict device installation or driver updates, causing “unknown device” behavior.

2) The “3-Layer” Troubleshooting Model (Firmware → Windows → Runtime)

Stop treating this like one problem. Treat it like three, and test each layer:

  1. Firmware/BIOS layer: Is the NPU exposed to the OS?
  2. Windows device layer: Does Device Manager show the correct NPU device + driver?
  3. Runtime/app layer: Does ONNX Runtime / DirectML / vendor EP see the NPU, and does your model actually compile for it?

3) BIOS/UEFI Switches That Commonly Hide the NPU (Snapdragon X / Ryzen AI / Core Ultra)

After BIOS updates, factory resets, or “Load Optimized Defaults,” AI-related toggles can revert. Names vary by OEM (Dell/HP/Lenovo/ASUS/Acer), but look for options like:

BIOS Settings Checklist (What to Look For)

  • AI / NPU / Neural Processing: Enable (may be under Advanced → CPU Features or Advanced → SoC).
  • UMA / Shared Memory (especially on AMD/Qualcomm platforms): Ensure shared memory isn’t constrained; some NPUs rely on shared system memory behavior.
  • Memory Integrity / VBS / Device Guard: Usually safe, but some early driver stacks fail under certain hardened settings. If you suspect this, test by temporarily disabling (advanced users only).
  • S3/S0ix / Modern Standby knobs: Rare, but power-state misconfigurations can break device enumeration after resume.
  • Discrete GPU priority / Hybrid graphics: Not a direct NPU toggle, but can influence DirectML/AI app backend selection (GPU takes priority; app never tries NPU).

Pro move: If the NPU vanished after a BIOS update, update to the newest BIOS again (not rollback), then re-check the AI/NPU toggles. Many OEMs quietly patch device enumeration bugs in later BIOS builds.

4) Verify NPU Presence in Windows (Device Manager + System Tools)

A) Device Manager

  1. Open Device Manager.
  2. Look under sections like System devices, Processors, or an NPU/Neural category (OEM-dependent).
  3. If you see Unknown device or Base System Device, that’s typically a missing OEM driver package.

Driver details check: Right-click the suspected device → Properties → Driver. If the provider is “Microsoft” with a generic date and your NPU is not recognized by AI apps, install OEM drivers (next section).

B) Event Viewer (When the Device Exists but Fails to Start)

Open Event Viewer → Windows Logs → System and filter for:

  • Kernel-PnP (device install/enum)
  • WHEA-Logger (hardware errors)
  • DriverFrameworks-UserMode (UMDF device failures)

Errors like “Device not started” or “Driver could not be loaded” often point to a mismatched driver version after Windows Update.

5) Install the Right NPU Driver Stack (OEM First, Then Vendor)

In 2026, the most reliable path is OEM support page → your exact model. Copilot+ machines often ship with tuned AI stacks, and Windows Update can replace them with less capable generic builds.

Best Practice Order

  1. Install latest chipset/SoC drivers (this can be prerequisite for NPU enumeration).
  2. Install the OEM NPU/AI driver package (or “AI Framework,” “Neural Processing,” “Qualcomm/AMD/Intel AI Runtime”).
  3. Reboot (don’t skip).
  4. Run Windows Update after OEM packages, not before—then re-check that Windows Update didn’t overwrite the NPU driver.

When to Use DDU / Clean Driver Removal

If you installed multiple AI runtimes and now your apps behave inconsistently (CPU-only one day, GPU-only the next), do a cleanup:

  • Uninstall OEM AI framework entries from Apps → Installed apps.
  • Remove the device in Device Manager (Uninstall device and check Delete the driver software if available).
  • Reinstall the OEM stack cleanly.

Why the enclosure recommendation? If you’re about to do driver surgery, having a fast external NVMe enclosure makes it painless to back up models, environments, and checkpoints—especially if you end up doing an in-place repair install of Windows 11.

6) Fix the Runtime Layer: ONNX Runtime, DirectML, and Vendor Execution Providers

Many “NPU not detected” reports are not about Windows failing to see the hardware—they’re about the AI runtime failing to select an NPU execution provider (EP).

A) Understand the Common Backends

  • CPU EP: Always available, slowest for AI.
  • DirectML (DML): Typically targets GPU first; some apps expose it as “Windows ML/DirectML.” DML doesn’t automatically mean “NPU.”
  • Vendor NPU EPs: Platform-specific paths (Qualcomm/AMD/Intel). These are where many apps fall short in 2026.

B) Update ONNX Runtime the “App-Compatible” Way

Power-user pitfall: updating ONNX Runtime system-wide can break apps that bundle a specific ORT build.

  • If your app bundles ORT: Update the app first; don’t override its runtime unless the developer documents it.
  • If you control the environment (Python): Pin versions. A known-good combo (example pattern) is often better than “latest everything.”

Recommended workflow (Python users): Create a fresh venv/conda env, install ORT + any vendor packages required by your platform, then test with a known NPU-friendly model.

C) Model Compatibility: The Silent NPU Killer

Your NPU can be fully functional and still refuse a model. Common blockers:

  • Unsupported ops / opset mismatch
  • Dynamic shapes (some NPUs want static shapes for compilation)
  • Quantization format differences (INT8 per-tensor vs per-channel, activation quantization assumptions)
  • Too much VRAM/RAM pressure causing fallback to CPU

Actionable test: Try a smaller, well-known ONNX model first (image classification or small embedding model). If that works on NPU, your problem is your model/export settings—not detection.

7) Windows Features and Settings That Affect Local AI Apps

A) Power Plan / Performance Modes

NPUs are power-efficient, but OEM “Silent” modes can throttle sustained AI workloads or delay device wake.

  • Set OEM utility to Balanced or Performance while testing.
  • Plug in power for initial benchmarking—some systems reduce accelerator clocks on battery.

B) Hyper-V, Core Isolation, and Sandboxing

Security features generally shouldn’t hide the NPU, but they can interfere with older driver stacks and some low-level runtimes. If you’re troubleshooting a stubborn machine:

  • Temporarily disable Memory integrity (Core isolation) for testing.
  • Reboot, test NPU detection, then re-enable if it wasn’t the cause.

Note: If you’re on a managed device, coordinate with IT—don’t bypass policy.

8) App-by-App Fixes: Why One AI App Sees the NPU and Another Doesn’t

In 2026, “NPU support” is still fragmented. Use this triage approach:

Step 1: Check the App’s Backend Menu

  • If the app only offers CPU and CUDA, it’s not using your NPU.
  • If the app offers DirectML, it may still be GPU-only depending on implementation.
  • Look for explicit options like NPU, QNN, Intel NPU, Ryzen AI, or Windows ML (NPU).

Step 2: Confirm the App’s Model Format Matches Its NPU Path

  • Some apps only run ONNX on NPU but use GGUF/Safetensors on CPU/GPU.
  • Some require a special export (static shapes, specific opset).

Step 3: Watch Real Utilization

Don’t trust “NPU enabled” checkboxes. Verify in Task Manager (Performance tab) if your Windows build exposes an NPU graph, or use OEM monitoring tools. If utilization stays at 0% while CPU spikes, you’re still on CPU fallback.

A USB-C power meter is a practical way to validate whether your workload moved from CPU/GPU to NPU: sustained wattage often drops when acceleration is truly happening.

9) Last-Resort Repairs (When Everything Looks Right but NPU Still Won’t Work)

A) In-Place Repair Install of Windows 11

If Device Manager shows the NPU but runtimes are broken across multiple apps, an in-place repair can reset corrupted system components without wiping data.

  • Download the latest Windows 11 installer (official).
  • Run setup → choose Keep personal files and apps.
  • Afterward, reinstall OEM chipset + AI/NPU packages.

B) Roll Back a Problematic Driver Update

If the NPU disappeared immediately after Patch Tuesday or an OEM updater run:

  • Device Manager → NPU device → Driver → Roll Back Driver (if available).
  • Then use OEM support drivers and pause driver updates (temporarily) while you stabilize.

Recommended Tools & Hardware (2026) for Local AI Troubleshooting

  • External NVMe SSD + enclosure: for backups, model storage, portable environments.
  • USB-C power meter: quick sanity check for actual accelerator efficiency gains.
  • Cooling pad (thin-and-light Copilot+ PCs): keeps boost behavior consistent during longer AI runs.

Explore More

FAQ

Why does Windows 11 show a Copilot+ PC, but my AI app says NPU not detected?

Because “Copilot+” is hardware branding, while your app needs a compatible runtime (ONNX Runtime/DirectML/vendor EP) and a model it can compile for NPU. Many apps still default to CPU/GPU unless explicitly configured.

Should I install NPU drivers from Windows Update or my laptop manufacturer?

Start with the manufacturer (OEM) package for your exact model, then apply Windows Update. OEM stacks often include paired components (chipset + AI runtime + firmware interfaces) that generic drivers may not.

DirectML is available—does that mean I’m using the NPU?

Not necessarily. DirectML commonly targets the GPU. Some apps expose DirectML but do not route workloads to NPU. Verify with utilization monitoring and real performance-per-watt changes.

My NPU shows in Device Manager. Why do models still fall back to CPU?

Usually model incompatibility (unsupported ops, opset, dynamic shapes, or quantization format), missing execution provider libraries, or the app packaging an older runtime. Test with a smaller known-good ONNX model in a clean environment.

Will enabling Hyper-V or Memory Integrity disable my NPU?

Typically no, but certain driver versions and runtimes can misbehave under hardened security. For troubleshooting, you can temporarily toggle these settings to isolate the cause, then re-enable them.