What Uses Your NPU on Windows: The App-by-App List

A flat NPU graph in Task Manager is the normal state of a Copilot+ PC, not a fault. The list of things that genuinely offload to the NPU on Windows is short and almost entirely first-party: Windows Studio Effects (camera blur, eye contact, auto framing, Voice Focus), Recall's on-device OCR, Live Captions translation — but not plain Live Captions — and the Microsoft-supplied Windows AI APIs, which Microsoft's own documentation says "always run on the NPU" on a Copilot+ PC. Third-party apps reach the NPU only through a vendor toolchain: OpenVINO on Intel, the QNN execution provider on Qualcomm, the Vitis AI execution provider on AMD. Ollama, LM Studio and llama.cpp do not use it at all. The full table is below, with the vendor gate and a verification step for every row.

The reason this is confusing is that Windows advertises the NPU as a capability and then gives you no way to see what is using it beyond a single utilisation graph. So people buy a 45- or 80-TOPS machine, open Task Manager, watch a flat line, and reasonably conclude something is broken. Usually nothing is. The features that use it are event-driven and brief — the NPU spikes while a camera effect initialises or a caption is translated, then goes quiet.

This page is a reference table, not an argument. Every row is sourced to a vendor or app-maker statement, and where a vendor claims acceleration without shipping anything you can independently confirm, it says so instead of repeating the claim.

What Actually Runs on the NPU? The Full Table

Yes means a first-party source states the work runs on the NPU. Partial means it depends on the silicon tier or on a plugin you install yourself. No means it does not, whatever the marketing suggests.

The shape of that table is the finding. Nearly everything in the "yes" column is Microsoft's own code. Windows Studio Effects and the Windows AI APIs are Microsoft-built models compiled for the NPU and shipped with the OS. Outside that set, an app has to be deliberately ported to one vendor's toolchain, and very few have been.

Why Is My NPU at 0% in Task Manager?

Because almost nothing is asking it to do work, which is the expected state.

Microsoft's Copilot+ PCs developer guide is direct about why: "Like other hardware resources on a PC, the NPU needs software to be specifically programmed to take advantage of the benefits it offers." An NPU is not like a CPU, which everything uses by default. It is closer to a codec block — dedicated silicon that sits idle until a piece of software is written against it.

Add to that the fact that the NPU workloads Windows actually ships are short. Background blur runs continuously but at very low utilisation, by design — the whole point is that it costs almost no power. Caption translation and OCR run in bursts measured in milliseconds. Neither produces the sustained graph you are used to seeing from a GPU under load.

The one case where 0% is genuinely a problem: if Settings → Bluetooth & devices → Cameras shows only "Basic settings" for your built-in camera on a machine that should support Studio Effects, the OEM's Studio Effects driver is missing or not installed. Microsoft's requirement is that the "hardware manufacturer must have opted in and installed the Windows Studio Effect driver on the system". That is an OEM driver package, so the fix is the manufacturer's update utility, not Windows Update.

Does Live Captions Use the NPU on a Ryzen 7040?

No, and there is no version of Live Captions that will, on that chip. This is the single most-asked version of the question, so it is worth separating the two features cleanly.

Microsoft's own support page splits it: "Live captions is available in Windows 11, version 22H2 and later", while "The ability to translate is available on Copilot+ PCs running Windows 11, version 24H2 and later."

So:

• Plain Live Captions (transcribing audio into captions on screen) predates Copilot+ entirely. It runs on any Windows 11 22H2 machine, including ones with no NPU. On your Ryzen 7040 it works, and it is not using the NPU to do it.
• Live Captions translation (Microsoft's page describes 40+ languages into English and 27 into Simplified Chinese) is Copilot+ only.

And a Ryzen 7040 "Phoenix" is not a Copilot+ machine. Its XDNA NPU is rated by AMD at 10 TOPS, against Microsoft's 40-TOPS bar. The Ryzen 8040 "Hawk Point" refresh is rated 16. Both are real NPUs, both run Studio Effects perfectly well where the OEM shipped the driver, and neither will ever get the Copilot+ feature set — the gate is a hardware rating, not a software flag someone can flip.

The AMD chips that do clear the bar are Ryzen AI 300 and AI PRO 300 (XDNA 2, rated 50 and 55 TOPS), Ryzen AI Max / Max+ 300, and Ryzen AI 400. Which line yours is in is settled by the exact part number — the full ratings table is on our every CPU with an NPU page.

Intel GNA vs the NPU: Are They the Same Thing?

No. Different block, different era, different toolchain — and only one of them appears in Task Manager.

• GNA — the Gaussian & Neural Accelerator — is a small, low-power block that shipped inside Intel Core processors before Core Ultra existed. It was built for lightweight audio and speech work: noise suppression, wake-word style detection, the kind of thing that has to run continuously without waking the CPU. It was never surfaced to users as an "NPU", never had a Task Manager tab, and was never a general-purpose inference target.
• The NPU, marketed by Intel as Intel AI Boost, arrives with Core Ultra (Meteor Lake) and is a substantially larger, general inference engine addressable through OpenVINO and Windows ML. Intel publishes a per-SKU "NPU Peak TOPS (INT8)" field for it on Intel ARK; it publishes no equivalent for GNA.

The practical consequence: if your machine predates Core Ultra, you do not have an NPU in the sense Windows means. A 12th- or 13th-Gen Core laptop with GNA cannot run Copilot+ features and will not show an NPU tab. It is not a driver problem and there is nothing to enable.

Within Core Ultra, the generations are far apart in rating too — Meteor Lake's NPU 3 is rated 11 TOPS and Arrow Lake's 13, both under the Copilot+ bar, while Lunar Lake's NPU 4 is rated 40 to 48 depending on SKU. Same "Core Ultra" brand, opposite sides of the gate.

Which Models Run on an NPU?

Not the ones in your Ollama library. NPU inference needs a model converted and quantised for a specific vendor's runtime, and Microsoft's documentation is explicit about why: "Many NPU devices, however, only support integer math in lower bit format, such as INT8, for increased performance and power efficiency. Therefore, AI models need to be converted (or 'quantized') to run on the NPU."

Where the ready-made ones live:

The mechanism that makes this work without you choosing anything is Windows ML. Microsoft describes the flow: Windows ML "queries the system for available hardware accelerators", selects the most performant execution provider "(for example, QNN for Qualcomm NPUs, OpenVINO for Intel NPUs)", loads it, and "if the preferred EP fails or is unavailable, Windows ML gracefully falls back to another (for example, using the GPU or CPU)".

That fallback is the reason a lot of "my NPU is not being used" reports are correct without anything being broken. The app asked for the best available accelerator, Windows ML could not place the model on the NPU, and it quietly ran on the GPU or CPU instead. Nothing errors. Nothing tells you.

For actually running a general local LLM, the NPU is still not the path — that argument, with the bandwidth arithmetic behind it, is in Copilot+ PC vs RTX for local AI, and if you are choosing between runtimes for a normal CPU or GPU setup, Jan vs LM Studio vs Ollama is the comparison you want instead.

How Do I Confirm Something Is Really Using the NPU?

Three levels, cheapest first.

1. The Task Manager graph. Open Task Manager → Performance. Microsoft's documentation confirms the tab exists: "For devices with NPUs, the Task Manager can now be used to view NPU resource usage", showing "the percentage of utilization, available memory, shared memory, driver version, physical location, and more". Trigger the feature you are testing — turn background blur on, run a caption translation — and watch for a spike. This tells you something used it, not what.

2. Trigger-and-toggle. For camera effects specifically, the cleanest test is a comparison: open Camera Settings, turn Studio Effects blur off, watch the graph while the camera runs, turn it on, watch again. If the NPU line moves with the toggle, that effect is on the NPU on your machine.

3. A trace, when you need to know which process. Microsoft ships a Neural Processing profile in Windows Performance Recorder. Their own quickstart, verbatim from the developer guide:

wpr -start ort.wprp -start etw_provider.wprp -start NeuralProcessing -start CPU
echo Repro the issue allowing ONNX to run
wpr -stop onnx_NPU.etl -compress

Open the resulting onnx_NPU.etl in Windows Performance Analyzer (free in the Microsoft Store) and expand "Neural Processing → NPU Utilization". WPR and WPA are part of the Windows Performance Toolkit, which comes with the Windows ADK. This is the only method that names the process submitting the work, so it is the one to use when a vendor claims acceleration and you want to check.

FAQ

Is my NPU broken if Task Manager shows 0%?

Almost certainly not. NPU work on Windows is event-driven and low-intensity by design — background blur is meant to be nearly invisible in the utilisation graph, and caption translation or OCR runs in bursts. The one real fault to check for is a missing OEM driver: if Settings → Bluetooth & devices → Cameras shows only "Basic settings" for your built-in camera, the manufacturer's Windows Studio Effects driver is not installed, and Windows Update will not supply it — the OEM's own update utility will.

Can I force an app to use the NPU?

Generally no, and there is no Windows-wide switch. Windows ML picks the execution provider automatically and falls back to GPU or CPU when it cannot place a model on the NPU. Where you do get a choice is inside apps built on a vendor toolchain — the OpenVINO plugins for Audacity, for instance, let you select the device in the effect dialog. Outside those, device selection is the app developer's decision, not a user setting.

Does Windows Studio Effects need a Copilot+ PC?

No, and this is the most useful thing on this page for owners of older AI laptops. Studio Effects requires Windows 11 22H2 or newer and an NPU the OEM opted in and shipped a driver for — there is no 40-TOPS gate. A Ryzen 7040 or a Meteor Lake Core Ultra 100 runs it. What the lower-capacity NPUs miss is the heavier variants: Microsoft notes that they "may support 'Standard Blur', but not 'Portrait Blur' due to compute limitations", and the same split applies to Teleprompter Eye Contact.

Does Zoom use my NPU for background blur?

If you enable blur in Windows Camera Settings rather than inside Zoom, yes — and Zoom does not need to know about it. Studio Effects is chained onto the end of the camera pipeline, so Microsoft's documentation states that once an effect is on, "it is on by default for any app using the camera, even if the app doesn't know about the effect". The trap is doing both: leave Zoom's own blur on as well and you get either a confused UI or two blur passes stacked, wasting exactly the power the NPU was supposed to save.

Which NPU specs matter for on-device AI?

Fewer than the marketing implies. The 40-TOPS Copilot+ threshold is a hard gate on the Windows feature set, so it matters as a yes/no. Above it, TOPS mostly stops differentiating, because every currently-shipping Windows part clears it. What actually decides whether an NPU is useful to you is whether a runtime exists that targets it for the work you want to do — OpenVINO, QNN or Vitis AI — and, for anything language-model-shaped, the machine's memory capacity and bandwidth rather than the NPU at all. If you are still choosing a machine, our NPU laptop table ranked by memory ceiling is built on exactly that logic.

Do Adobe or DaVinci Resolve use the NPU?

We could not source a first-party statement from either company that a shipping feature targets the Windows NPU specifically, rather than the GPU, so this page does not claim one. Both ship features described as AI-accelerated; on Windows those are overwhelmingly GPU workloads. If you see a claim otherwise, check whether it names the NPU explicitly or just says "AI acceleration" — and if it matters to your purchase, the WPR trace above will settle it on your own machine in five minutes.

What This Page Does Not Claim

• Nothing here was benchmarked. Every row is a vendor or app-maker statement, quoted and linked. We have not measured NPU utilisation on any machine and do not publish numbers we did not measure.
• The Windows AI API list moves. Microsoft has stated that Phi Silica "is being replaced by Aion Instruct", rolling out to Insider builds in October 2026 and retail in November 2026. Check the current API table before building anything on it.
• Third-party support is where this page will age. Vendor toolchains are adding applications steadily; the "no" and "not established" rows are a snapshot, not a permanent verdict. The verification method in the section above outlives the table.
• OEM driver behaviour varies. Whether Studio Effects appears at all on a given laptop is the manufacturer's decision, not Microsoft's, so two machines with the same chip can behave differently.

Sources

• Microsoft — Copilot+ PCs developer guide (40-TOPS threshold, Task Manager NPU reporting, Windows ML execution-provider selection, WPR tracing quickstart, model quantisation requirement)
• Microsoft — What are Windows AI APIs? (the per-API NPU / GPU / CPU support table, model preinstall and download behaviour, Phi Silica to Aion Instruct transition)
• Microsoft — Windows Studio Effects overview (effect list, driver and OEM opt-in requirement, lower-capacity NPU limitations, camera-pipeline behaviour, the Teams background-segmentation note)
• Microsoft — Retrace your steps with Recall (Secured-core, 40 TOPS, 16 GB RAM, 256 GB storage requirements)
• Microsoft — Use live captions to better understand audio (captions on 22H2+, translation on Copilot+ 24H2+)
• Qualcomm AI Hub — Compute models (the pre-converted model catalogue and runtime options)
• AMD Ryzen AI Software documentation (Vitis AI Execution Provider, OGA and llama.cpp paths)
• intel/openvino-plugins-ai-audacity (effect list and the CPU/GPU/NPU device statement)
• Intel ARK (per-SKU NPU Peak TOPS (INT8) fields)