AI PC TOPS Explained: Why NPU Ratings Don’t Predict Copilot+ Performance

The argument holds up, but a few of its claims need checking. Here's what the manufacturers' own documents support, what they don't, and what Windows buyers and IT departments should look at besides the sticker.

The 40 TOPS line is Microsoft's​

The most important NPU number in Windows comes from Microsoft. Its Copilot+ PC business page describes the category as Windows 11 devices whose NPUs can run more than 40 TOPS. It says these NPUs sit alongside the CPU and GPU and take on some AI tasks so the other processors have more room.

The threshold controls a specific set of features, not AI on Windows in general. Microsoft lists Recall (preview), Click to Do, Live Captions with translation, Windows Studio Effects and improved Windows search as Copilot+ features. It also says these experiences vary by device and region and that some are still rolling out through updates.

The details show that 40 TOPS is required but doesn't guarantee a feature:

• Recall requires a Copilot+ PC, at least 256 GB of storage and at least 50 GB free. It also needs Windows Hello Enhanced Sign-in Security and the April 2025 non-security preview update or later. Microsoft says Arm64EC models aren't supported.
• Windows Studio Effects on Microsoft Learn requires Windows 11 version 22H2 or newer, a supported NPU, and a Studio Effects driver that the laptop maker has chosen to install. The developer documentation says lower-capacity NPUs may support Standard Blur but not Portrait Blur, and may support Standard Eye Contact but not the Teleprompter version.
• Live Captions runs on ordinary Windows 11 PCs. Real-time speech translation is the part Microsoft reserves for Copilot+ machines.

Quick check: To see whether your laptop has Studio Effects, look for "Studio effects" in Quick Settings. You can also go to Settings > Bluetooth & devices > Cameras, choose your camera, and look for a "Windows Studio Effects" section under the preview. Microsoft's documentation says that if you see only "Basic settings," the laptop probably doesn't support Studio Effects.

Section summary: 40 TOPS is the entry requirement for Copilot+ features. Each feature also has its own conditions for drivers, storage, sign-in hardware and Windows updates.

Bigger numbers, more model variants​

NPU ratings have climbed quickly. Qualcomm's September 2025 announcement said the Hexagon NPU features 80 TOPS of AI processing—the world's fastest NPU for laptops—to enable concurrent AI experiences in Copilot+ PCs. That's up from 45 TOPS on the first Snapdragon X Elite. Reporting on the launch noted that the NPU is rated for 80 TOPS (INT8) on the chip, which far outstrips the 50-to-55-TOPS-rated NPUs to date from AMD and Intel for mobile computers.

The family name tells you less than it seems. Qualcomm's current Snapdragon X2 Elite product page splits the lineup by exact part number. It lists X2E-96-100, X2E-90-100, X2E-84-100 up to 85 TOPS and X2E-94-100, X2E-88-100, X2E-80-100, X2E-78-100 80 TOPS. Qualcomm's own X2 Elite product brief, meanwhile, shows the flagship X2E-96-100 at 80 TOPS. So the same chip carries two different ratings in two official documents, and neither one is a typo on a retailer's site. That supports the Vocal essay's main advice: check the full part number and the vendor's current specification page, not the marketing name.

Qualcomm's lineup also differs in ways that have nothing to do with NPU compute. In the product brief, the top Extreme parts pair LPDDR5x memory with 228 GB/s of bandwidth. A retailer's review says standard X2 Elite models support up to 128GB LPDDR5x with bandwidth reaching 152GB/s. For AI workloads that are limited by memory bandwidth, that gap can matter as much as a few TOPS.

The other vendors in the essay:

The essay also says AMD sells same-year chips rated at both 50 and 60 TOPS under Ryzen AI branding. The AMD releases in the evidence gathered here don't confirm that for the Ryzen AI 300 series. Treat it as a reason to check each model's spec sheet, not as an established fact.

Section summary: The number on the box describes one specific chip, and sometimes one specific version of a document. Match the exact processor model to the current vendor spec before comparing laptops.

Why two "50 TOPS" chips may not match​

TOPS doesn't state the workload, how long it runs, the software path or the numerical precision. Lower precision fits more operations into the same silicon, so the precision behind a headline number matters. Qualcomm labels its X2 figures as INT8. Apple's M4 announcement describes the Neural Engine as capable of "38 trillion operations per second" and gives no precision. That makes a direct comparison with an INT8-rated Windows NPU less reliable than a spec table suggests.

The essay's most concrete example comes from independent research on Apple's Neural Engine. A public GitHub project, maderix/ANE, uses reverse-engineered private Apple APIs to run workloads directly on the M4's Neural Engine. The project reports peak FP16 throughput of about 18.6 TOPS on a particular convolution benchmark and 35.1 TOPS with INT8 weights and activations. That's a 1.85x to 1.88x speedup on those specific configurations. The author credits much of the gain to INT8 activations halving the SRAM bandwidth between tiles, not to faster math.

Two caveats matter here:

1. This is research, not a benchmark suite. The project calls itself a proof of concept built on private, undocumented APIs that could break with any macOS update. It says its training runs use only about 5-9% of peak.
2. It doesn't show that Apple's rating is false. It shows that on the tested operations, lower precision helped mainly by moving less data. That supports the essay's broader point: data movement, not raw compute, often sets the limit.

Section summary: Precision, workload and software all affect a TOPS figure. Equal numbers from different vendors don't guarantee equal real-world results.

The data-movement argument, and where it's the author's reading​

The essay starts with Intel's Lunar Lake launch. Intel's NPU material lists up to 48 TOPS, and the essay cites Intel's product brief for "up to 2x bandwidth" compared with the previous generation. Intel also moved system memory onto the processor package. Valiani argues that together these show a chipmaker spending real money on moving data, not only on faster math.

He is open that this is his interpretation. Intel's public reasons for on-package memory focus on lower power and a smaller motherboard, not on feeding the NPU. The same caution applies to a claim he cites about Intel doubling the NPU's DMA engines, which move data inside the chip. That claim comes from press accounts of a technical briefing, and an independent tester reportedly measured the DMA hardware as weaker than the previous generation's. Valiani says he couldn't reconcile the two. Neither claim was independently verified here, so treat them as open questions.

The broader engineering principle is well established. For low-batch work, such as a local chatbot producing one token at a time, a fast compute engine often sits idle waiting for data from memory. That's why the memory bandwidth figures in Qualcomm's product tables deserve as much attention as the NPU line.

Heat: the spec nobody prints​

A peak rating describes a chip at its best. A thin laptop has one power and thermal budget shared by the CPU, GPU and NPU. The essay points to a March 2026 arXiv study of on-device language-model inference. Its abstract reports that an iPhone 16 Pro lost nearly half its throughput within two iterations. Valiani adds that this test ran on the phone's GPU, not the Neural Engine. That detail comes from his reading of the paper and wasn't independently confirmed here. Either way, a phone test says nothing specific about laptop NPUs, and no equivalent public sustained-load results were found for the laptop chips discussed.

Commentary on the Snapdragon X2 launch made a similar point. Analysts noted that performance depends on OEM thermals, software support, and workloads, and one retailer's review put it bluntly: "TOPS" measures theoretical speed limits, not how smoothly your everyday tools will actually run.

The essay also asks whether Microsoft's Copilot+ certification tests sustained NPU performance or only checks the rated peak. Microsoft's public pages don't answer that.

A practical buying checklist​

Individual buyers and IT departments ordering in bulk can use the same steps:

1. Start with the feature. Decide whether you need Recall, live translation, Studio Effects or semantic search, and check that feature's requirements, not just the Copilot+ label.
2. Check the full processor part number. "Snapdragon X2 Elite" or "Core Ultra 200V" names a family, not one chip.
3. Confirm the figure is NPU-only. Some marketing adds CPU and GPU compute into a combined "platform TOPS" number.
4. Check memory bandwidth and capacity. Local AI models are often limited by memory, and Recall alone needs 50 GB free.
5. Check drivers and updates. Studio Effects depends on the laptop maker's driver, and Copilot+ features keep arriving through Windows updates.
6. Look for sustained, task-specific testing. Independent reviews of the exact laptop running your workload are worth more than any peak rating.
7. For fleets, standardize on a SKU. When you order a hundred "identical" machines, ask for the exact processor and memory configuration in the contract.

The bottom line​

TOPS isn't a fake number. Microsoft's 40 TOPS threshold is a real engineering requirement, and every major Windows chip vendor has models that meet it. But TOPS is a peak rating quoted at a stated or unstated precision, and it tells you nothing about memory bandwidth, software support or how the laptop behaves after ten minutes of heat. As NPU ratings keep rising, those factors will increasingly decide how much of the headline number reaches real work, and none of them appear on the box.

For WindowsForum readers, related threads on Copilot+ PC feature rollouts, Recall privacy controls, Windows on Arm app compatibility and enterprise laptop standardization are good places to continue.