NPUs and Edge AI: What Chip Vendors Actually Document

What an NPU is, in vendor words

A neural processing unit is, in Apple's published definition from its M4 announcement, "an IP block in the chip dedicated to the acceleration of AI workloads." That definition holds across vendors: an NPU is a fixed-function accelerator designed to run the arithmetic of neural networks — especially inference — more efficiently per watt than a general-purpose CPU or GPU.

The efficiency argument is the vendors' own. Intel documents its fourth-generation NPU (NPU 4.0) in Lunar Lake as "ideal for running sustained AI workloads while remaining energy-efficient." AMD documents its XDNA 2 architecture as "designed to run models continuously at the lowest power levels possible," claiming up to 2× the power efficiency of its previous generation. The pitch is consistent: offload always-on AI features from CPU and GPU to a block that sips power, and battery life survives the AI era.

TOPS: what the number does and does not mean

TOPS — trillions (or tera) operations per second — is the industry's headline metric for NPU throughput. It counts how many arithmetic operations the accelerator can theoretically perform per second, at a given numerical precision (most often INT8). AMD's own glossary is unusually candid about the limits of the figure: "Trillions of Operations per Second (TOPS) for an AMD Ryzen processor is the maximum number of operations per second that can be executed in an optimal scenario and may not be typical. TOPS may vary based on several factors, including the specific system configuration, AI model, and software version."

Two structural ambiguities matter when comparing spec sheets. First, precision: an INT8 TOPS figure is not comparable to an FP16 figure, and vendors do not always state which they are quoting. Second, scope: Intel documents "up to 120 total platform TOPS" for its Core Ultra 200V series — but explicitly counts that as spread "across central processing unit (CPU), graphic processing unit (GPU) and neural processing unit (NPU)." The NPU alone in the same platform is documented at "up to 48 TOPs." A buyer comparing "120 TOPS" against a competitor's "50 TOPS NPU" figure is comparing different things entirely.

Microsoft's 40+ TOPS line in the sand

The number that reshaped the market is Microsoft's, not any chipmaker's. Microsoft's published Windows 11 specifications state that Copilot+ PC features "require a neural processing unit (NPU) capable of 40+ trillion operations per second (TOPS) to function as intended." Its developer documentation repeats the threshold: "Many of the new Windows AI features require an NPU with the ability to run at 40+ TOPS."

Microsoft also publishes the rest of the Copilot+ class definition. Its business-facing documentation describes Copilot+ PCs as having "a neural processing unit (NPU) capable of over 40 trillion operations per second (TOPS)," requires "a minimum of 256 GB of hard drive capacity, with at least 50 GB free for Recall to operate," and documents that Recall is allocated 25 GB by default on a 256 GB PC — roughly three months of activity, per Microsoft's own text. Sample configurations Microsoft lists pair 16 GB of memory with 256 GB of storage.

The documented numbers: Qualcomm, AMD, Intel, Apple

Each vendor documents its NPU differently — and the differences in how they publish are as telling as the numbers.

Apple is the notable case: at 38 trillion operations per second it sits below Microsoft's 40+ TOPS Copilot+ threshold, and Apple does not use "TOPS" branding for the Neural Engine on the fetched page — it says "trillion operations per second" without stating the precision. Apple's claim that the M4 Neural Engine was "faster than the neural processing unit of any AI PC today" was made at its May 2024 announcement and predates the current generation of 45–55 TOPS Windows silicon; vendor superlatives decay quickly in this market.

AMD's progression is also documented on its own pages: the same partner article notes the XDNA 2 generation delivers "up to 3X more AI performance than previous generation AMD products." Third-party coverage traces the ladder (10 TOPS in Ryzen 7040, 16 TOPS in Ryzen 8040, 50–55 in AI 300) Not verified against AMD primary pages in this session — the prior-generation figures are not on the pages we fetched.

What Microsoft documents as running on the NPU

Microsoft publishes a feature list gated on the Copilot+ NPU threshold, with per-feature caveats that are worth reading closely. From its Windows 11 specifications page: Automatic super resolution (a curated set of games, Snapdragon X Series devices only, per Microsoft's note), Click to Do ("available image and text actions vary by device, region, language, and character set"), Paint Cocreator and Object select, Photos Image creator / Relight / Restyle / Super resolution (Relight listed for Snapdragon X Series devices), Snipping Tool "Perfect screenshot," Voice Access flexible commands, and Windows Studio Effects ("requires a compatible camera; available effects vary by device").

Several of these features Microsoft ties specifically to Snapdragon X Series hardware in the published notes — meaning a 45–55 TOPS AMD or Intel Copilot+ PC is not guaranteed every feature in the list at launch. Microsoft also documents that some actions require a Microsoft account, an internet connection, or a subscription. The NPU threshold unlocks the class; individual features carry their own device, region and rollout conditions.

The fine print vendors publish about their own claims

Vendors do document limits — usually in footnotes. AMD's TOPS glossary (quoted above) disclaims typicality and ties results to configuration, model and software. Intel footnotes its own TOPS figures as "'up to' and approximate until final IP frequency defined," noting that different SKUs with different frequency and power targets will deliver different TOPS. Microsoft's Copilot+ pages carry the standard "features and app availability may vary by region and market" and note that some features need "additional hardware."

Read together, the fine print says something no headline number does: TOPS is a ceiling measured under conditions the vendor chooses. It is a useful floor for comparison — a 15 TOPS chip will not join the Copilot+ class — but it is not a performance guarantee for any workload you actually run.

What no vendor documents

The gaps in published documentation are consistent across all four vendors. None publishes sustained (thermal-limited) NPU throughput for a real model — only peak figures. None publishes tokens-per-second for a named local LLM running on its NPU. None documents NPU memory bandwidth separately from system memory Not verified on the pages fetched here. And none publishes a cross-vendor benchmark methodology for its TOPS claims — each vendor's "up to 5X" comparisons reference its own test setups.

There is also a documentation asymmetry between marketing and engineering pages: Qualcomm's 45 TOPS figure lives in the product page's structured data; Intel's 48-TOPS NPU figure appears in a support overview article; AMD's 50/55-TOPS table sits in a partner-facing article rather than the consumer product page; Apple's figure appears in a newsroom announcement. Finding the number is half the work — which is precisely why this article collects the URLs.

Decision table: reading an AI PC spec sheet

The practical summary: an NPU is a real, documented efficiency win for always-on AI features, and Microsoft's 40+ TOPS threshold gives buyers one genuinely comparable line. Beyond that line, vendor numbers diverge in precision, scope and measurement conditions — and the fine print, which every vendor publishes, is where the honest comparison lives.