You installed UL Procyon AI Inference Benchmark, selected your Copilot+ NPU as the target, hit run — and got an error instead of a score: a backend initialization failure, a “device not available” message, a crash halfway through, or a result that quietly ran on the GPU. The CPU run works fine, which makes the NPU failure feel personal. It isn’t — it’s structural.

Procyon’s NPU path is the most demanding code path in the benchmark. The CPU and GPU paths ride on mature stacks, but the NPU path depends on a three-way handshake between Procyon’s NPU backend, the DirectML runtime, and your vendor’s NPU driver — and all three move fast. A Procyon build from six months ago may not understand a driver from last month. When the handshake fails, you get errors that point everywhere and nowhere.

This guide gives you six fixes in order: update Procyon itself, install the runtimes its NPU backend needs, repair the NPU driver, force explicit NPU device selection, quiet the services that steal the NPU mid-run, and finish with a clean reinstall if the installation state is corrupt.

Why this happens

  • Outdated Procyon build: NPU backends evolve with vendor runtimes — an old Procyon may not recognize a new NPU driver.
  • Missing DirectML or vendor runtime: The NPU backend needs runtime components the CPU path never touches.
  • Stale or generic NPU driver: A Windows Update-swapped driver exposes the NPU to Task Manager while hiding it from benchmark backends.
  • GPU-first device enumeration: Procyon can silently target the iGPU when the NPU isn’t explicitly selected.
  • Background services holding the NPU: Recall, camera effects, and audio pipelines can lock NPU resources mid-benchmark.
  • Corrupted install or blocked files: Antivirus quarantines and half-updated installs break the NPU backend’s DLL loading.

How to fix it

Work through these in order — diagnose first, then act.

Fix 1: Update Procyon to the latest build

Start with the app itself, because NPU backend support moves faster than any other part of Procyon. UL ships backend updates alongside vendor runtime releases — support for newer Snapdragon, Intel AI Boost, and Ryzen AI drivers lands in Procyon updates, not Windows Update. A Procyon build older than your NPU driver is a guaranteed mismatch: the benchmark doesn’t know how to talk to your driver version. This is the cheapest fix and resolves many “broke after a driver update” reports. Check the version inside the app, since Store and standalone installers can drift apart.

  1. Check the exact build number in Procyon’s About or settings screen.
  2. Update through your original install channel — don’t mix Store and standalone.
  3. Check UL’s release notes for NPU backend changes affecting your chip.
  4. Restart the laptop after updating so driver bindings refresh.
  5. Retry the NPU benchmark before touching anything else.

Fix 2: Install the runtimes the NPU backend needs

Procyon’s NPU backend isn’t self-contained — it leans on system runtimes the CPU path never loads. The DirectML runtime, the Visual C++ redistributables, and on some chips a vendor NPU runtime package must all be present and current; a missing one produces backend initialization errors that read like driver faults. Windows 11 ships a DirectML version, but Procyon’s backend may need a newer one than your OS build carries. Install the pieces deliberately instead of assuming the OS has them.

  1. Install the latest Windows 11 cumulative update for the current inbox DirectML.
  2. Install the latest Visual C++ Redistributable (x64, plus ARM64 on Snapdragon).
  3. Check your NPU vendor’s guidance for any extra runtime package.
  4. Restart after runtime installs — DirectML binds at boot.
  5. Retry the benchmark; a changed error message is progress.

Fix 3: Repair the NPU driver

If the backend still fails to initialize, suspect the classic Windows Update swap: a validated vendor NPU driver replaced by a generic one that shows the NPU in Task Manager but exposes nothing usable to DirectML clients like Procyon. The symptom is distinctive — Device Manager looks healthy, the CPU benchmark runs, but every NPU attempt fails at initialization. Reinstall from the authoritative source: the laptop maker’s package for Snapdragon, Intel’s driver detection tool for Intel AI Boost, or the OEM/AMD package for Ryzen AI.

  1. Confirm the NPU device is clean under Neural processors in Device Manager.
  2. Download the current NPU driver from the OEM or vendor source above.
  3. Install as administrator and restart when prompted.
  4. Confirm the NPU appears in Task Manager’s Performance tab.
  5. Retry the NPU run — errors now point back to Fix 2.

Fix 4: Force explicit NPU device selection

Procyon enumerates DirectML devices in system order, and on nearly every AI-PC laptop the integrated GPU comes before the NPU. If the run configuration doesn’t explicitly select the NPU backend, the benchmark happily runs on the iGPU — producing a suspicious result or an error when the expected NPU backend never engages. Don’t trust the default target: open the run settings, verify the device dropdown names the NPU specifically, and check the run log afterward for the actual execution device. A result with zero NPU utilization in Task Manager is a misconfigured run, not a failed NPU.

  1. Open Procyon’s benchmark settings before starting.
  2. Set the inference device/backend explicitly to the NPU — not Auto or GPU.
  3. Start the run with Task Manager’s Performance tab visible on the NPU graph.
  4. After the run, check Procyon’s log for the actual execution device.
  5. If the log shows the GPU, reconfigure and rerun.

Fix 5: Quiet the system services competing for the NPU

The NPU is shared, and several Windows features hold it during a benchmark. Recall’s background processing is the heaviest consumer — its disk and NPU activity mid-run steals cycles and injects variance, and the Recall high disk I/O fix covers quieting it. Next is the camera pipeline: Studio Effects and Windows Hello’s face pipeline keep NPU-backed camera processing alive, and machines with Windows Hello face issues on AI PCs often have that pipeline stuck in a state that blocks benchmark access. Even audio matters — apps holding Bluetooth LE Audio streams on ARM keep real-time DSP paths warm. Pause them all.

  1. Close every app except Procyon — browsers, Teams, camera apps, music players.
  2. Pause Recall snapshotting and Windows Search indexing for the run.
  3. Disconnect Bluetooth audio devices and close apps using the camera.
  4. Set Best performance power mode and plug in the charger.
  5. Rerun and compare — reduced variance confirms contention was the cause.

Fix 6: Clean-reinstall Procyon and run as administrator

When every layer is correct and the NPU backend still fails, the installation itself is suspect: half-applied updates, antivirus-quarantined backend DLLs, or stale config from an older build can poison the NPU path while CPU runs stay fine. A repair install rarely clears this — do a full uninstall, manually remove leftover folders, re-download the installer, and reinstall fresh. Running the first NPU benchmark as administrator also rules out device-access permission issues. If it still fails, report to UL with your driver versions — and check Microsoft’s Windows support pages for known DirectML issues on your OS build first.

  1. Uninstall Procyon fully; delete leftovers in Program Files and AppData.
  2. Add Procyon’s install folder to the antivirus exclusion list.
  3. Re-download the installer from your original channel and install fresh.
  4. Run the NPU benchmark as administrator.
  5. If it still fails, collect the error text, driver versions, and OS build for UL support.

Closing

Procyon NPU errors are handshake failures between the benchmark’s backend, the DirectML runtime, and the NPU driver — update Procyon first, install the runtimes second, repair the driver third, then select the NPU explicitly and quiet competing services. Work the handshake in that order and the score appears.

Frequently asked questions

Why does Procyon fail on the NPU but work on the CPU?
The CPU path needs almost nothing beyond the app itself, while the NPU path depends on the DirectML runtime, the vendor NPU driver, and the correct NPU backend build of Procyon. Any gap in that stack fails only the NPU run, which is why the CPU result looks fine.
Do I need a specific Procyon version for my NPU?
Yes. UL updates Procyon’s NPU backends as vendors ship new runtimes — an older Procyon build may not recognize a newer NPU driver, and vice versa. Keep Procyon updated to the latest build before troubleshooting anything else.
Can antivirus software break the Procyon NPU benchmark?
It can interfere, yes. Real-time scanning of Procyon’s model files and runtime DLLs sometimes blocks the NPU backend’s file access mid-run, producing errors that look like driver faults. Add Procyon’s install folder to the exclusion list if runs fail inconsistently.
Why does Procyon pick my GPU instead of the NPU?
Procyon enumerates DirectML devices in system order, and the integrated GPU usually comes first. If the NPU backend isn’t explicitly selected in the benchmark settings, the run silently targets the GPU — check the run configuration before assuming the NPU failed.