New Surface PCs Get 78% More NPU TOPS. What That Means for Buyers

On September 23, 2026, Microsoft announced a new Surface Pro 12-inch and Surface Laptop 13-inch built on Qualcomm's Snapdragon X2 Plus, a chip Qualcomm rates at 80 TOPS for its neural processing unit (NPU), with the devices on sale from October 13 in select markets. Microsoft's claim of "95% faster on-device AI" came from a benchmark, not from that number, and it exceeds the TOPS increase. For anyone shopping for an AI PC, that is the lesson: TOPS tells you whether a laptop clears Microsoft's Copilot+ bar, not how fast your AI tools will run.

Key takeaways

• TOPS (trillions of operations per second) is a peak math rate, usually quoted at 8-bit integer (INT8) precision; AMD's own footnote says it "may not be typical."
• Copilot+ features such as Recall and Click to Do need an NPU rated at 40+ TOPS, plus at least 16 GB of RAM and 256 GB of storage. A high "overall" or GPU TOPS figure does not count toward that bar.
• For local chatbots, memory bandwidth and RAM limit speed more than TOPS: an 8-billion-parameter 4-bit model has a ceiling of about 38 tokens (word pieces) per second on 152 GB/s, using under 1% of an 80 TOPS NPU.
• An NPU only helps when an app is built to use it. Popular local-AI tools such as Ollama list GPU acceleration only.
• Benchmarks that run real models (MLPerf Client, Procyon AI, Geekbench AI) are a better guide than spec-sheet TOPS.

What it means for you

• Buying a laptop for everyday work or school: any current Copilot+ laptop clears the bar; beyond 40 NPU TOPS, battery life, screen and RAM matter more than a bigger AI number.
• Wanting Copilot+ features: check for the Copilot+ PC label or an NPU rated 40+ TOPS, 16 GB of RAM, 256 GB of storage and Windows 11 version 24H2 or newer. All are required, and Recall adds security conditions of its own.
• Running local AI models or creative AI tools: prioritize RAM and memory bandwidth, then check which hardware your apps actually use.
• Gamers with a discrete GPU: a recent card typically outruns a laptop NPU on raw AI math but does not make a PC Copilot+, which requires an NPU.
• Keeping an older PC: a first-wave "AI PC" with an 11 TOPS NPU will not gain Copilot+ features through updates, but GPU-based and cloud AI tools still work.

What a TOPS number actually counts

An NPU is a processor block built for the matrix math inside AI models. TOPS counts the basic operations it can complete per second at full clock speed with every unit busy. Three details change what that means.

Precision: INT8 is the usual yardstick

Precision is how many bits represent each number. Intel's spec pages label NPU figures "NPU Peak TOPS (Int8)," and Qualcomm's Snapdragon X2 Plus product brief lists "TOPS (INT8) 80." The same hardware typically manages fewer operations per second at 16-bit floating point (FP16) and, where it supports 4-bit math, more at 4-bit, so figures at different precisions are not comparable. Microsoft's Copilot+ PC developer guide notes that many NPUs only support lower-bit integer math, so models must be converted ("quantized") to run there.

Sparsity can double the headline

Some chips can skip multiplications by zero. In NVIDIA's description of structured sparsity, at least two of every four values must be zero, and throughput is then "theoretically" 2x the dense rate. A sparse figure assumes a specially pruned model. Intel's spec pages mark "Sparsity Support: Yes" without saying whether the TOPS line includes it, so compare figures only when both state precision and sparsity.

Peak is not sustained

TOPS assumes full clock speed. A thin laptop on a long AI job may slow down to stay within power and heat limits; the new Surface Pro 12-inch is, in Microsoft's words, "fanless." AMD's CES 2026 press release calls TOPS the maximum "in an optimal scenario," varying with system configuration, AI model and software version.

NPU TOPS vs. "overall" TOPS

Vendors quote AI figures for different parts of a chip. The table uses each maker's own published numbers as of October 2026; each chip name links to its source.

Intel's "overall" figure adds up the chip's parts. For the 258V, 115 − 64 (GPU) − 47 (NPU) leaves 4 TOPS, which by elimination is the CPU. For the 155H, 33 overall sounds close to 40, but its NPU alone is 11.

Apple gives no precision for the M4's 38 trillion and no Neural Engine figure for the M5, so Apple numbers do not line up with Windows spec sheets. NVIDIA promotes FP4, a lower precision than INT8, next to its 3,352 figure without labeling it, so a straight ratio likely overstates the gap; a desktop card is still far faster at raw AI math and draws hundreds of watts doing it.

The Copilot+ line: 40 NPU TOPS

Microsoft's Copilot+ PC explainer (December 2025) sets the minimum as an NPU of at least 40 TOPS, 16 GB of RAM, 256 GB of SSD storage and Windows 11 version 24H2 or newer. The requirement names the NPU, so platform or GPU TOPS do not substitute for it. Features that depend on the NPU include Recall, improved Windows search, and Click to Do, described in Microsoft's April 2025 feature update, along with Live Captions translation, Cocreator in Paint, AI editing in Photos and Windows Studio Effects.

Rollout has varied by chip and region: Microsoft's April 2025 update brought Click to Do text actions to Snapdragon PCs first, with AMD and Intel models promised "in the next few months," and said Recall and Click to Do would reach the European Economic Area later in 2025. Recall also has conditions beyond the hardware: Microsoft's Recall support page requires a Copilot+ PC that meets Microsoft's Secured-core standard, at least 50 GB of free storage, Device Encryption or BitLocker, and Windows Hello Enhanced Sign-in Security with a biometric sign-in, and snapshot saving is off until you opt in.

Memory, not TOPS, paces chatbots

A chatbot answer has two phases, as NVIDIA's guide to LLM inference explains. First the model reads the whole prompt in one large parallel calculation that can keep the compute units busy; here TOPS helps, setting the wait before the first word. Then it writes one token at a time, rereading the model's weights from memory for each. NVIDIA calls that phase "memory-bound": moving data, not doing math, sets the speed.

That gives a simple ceiling for writing speed:

max tokens per second ≈ memory bandwidth (GB/s) ÷ model size in memory (GB)
model size (GB) ≈ parameters (billions) × bytes per parameter

Bytes per parameter is 2 at FP16, 1 at INT8 and 0.5 at 4-bit. Real speeds fall below the ceiling because other data and apps share the bandwidth.

Illustrative example. Take an 8-billion-parameter model on a laptop with 152 GB/s of bandwidth, the Snapdragon X2 Plus figure:

• FP16: 8 × 2 = 16 GB, so at most 152 ÷ 16 = 9.5 tokens per second, and it cannot fit alongside Windows in 16 GB of RAM.
• INT8: 8 GB, at most 19 tokens per second.
• 4-bit: 4 GB, at most 38 tokens per second. On the X2 Elite Extreme's 228 GB/s, the ceiling rises to 57.

Real 4-bit files run somewhat larger than 4 GB because they also store scaling data, so the true ceiling is a little lower.

Now the compute side. A transformer model needs roughly 2 operations per parameter per token, per the scaling-laws paper by Kaplan and colleagues: 16 billion operations per token here. At 38 tokens per second that is about 0.6 trillion operations per second, under 1% of an 80 TOPS NPU.

Doubling TOPS would barely change writing speed; doubling bandwidth or halving model size would roughly double it. TOPS shows up in the prompt phase: a 4,000-token document is about 64 trillion operations, 0.8 seconds at a perfect 80 TOPS versus 1.6 at 40, before real-world losses.

RAM capacity is the other limit: model, working memory and Windows must all fit. AMD pitches its Ryzen AI Halo system, with up to 128 GB of unified memory, for models of up to 200 billion parameters, though its Ryzen AI Max+ chips list a 50 TOPS NPU, below the mainstream HX 475's 60.

Software decides which chip works

An NPU sits idle unless the app sends work to it. Microsoft's developer guide says the NPU "needs software to be specifically programmed" to benefit.

Its recommended route, Windows ML, is Microsoft's maintained copy of the open-source ONNX Runtime and replaces DirectML as the recommended path. When Windows ML became generally available on September 23, 2025, it supported vendor plug-ins called execution providers: AMD's Vitis AI, Intel's OpenVINO, Qualcomm's QNN and NVIDIA's TensorRT for RTX. If the preferred one is unavailable, it falls back to the GPU or CPU, so an app can run without touching the NPU.

Tool support varies widely:

• Windows' own Copilot+ features are designed to run on the NPU.
• Ollama, a popular way to run chatbots locally, lists NVIDIA, AMD Radeon, Apple Metal and Vulkan GPU support and does not mention NPUs.
• AMD's Ryzen AI software offers NPU-only, hybrid NPU-plus-graphics and GPU-only (via llama.cpp) modes on Ryzen AI 300-series chips; older Ryzen 7000/8000 chips are limited to GPU and CPU for language models.
• Intel lists OpenVINO, Windows ML, WebNN and ONNX Runtime as the frameworks the Core Ultra X9 388H's NPU supports.

Benchmarks that test real work

Three benchmarks run actual models instead of quoting theoretical peaks.

• MLPerf Client from MLCommons measures time to first token (the prompt phase) and tokens per second (the writing phase). Version 2.0, released August 18, 2026, uses Phi 4 Mini Instruct, adds experimental Qwen 3 8B and image-generation tests, and times agent-style tasks end to end.
• UL Procyon AI Computer Vision times image classification, captioning, object detection, segmentation and upscaling models on a chosen GPU or NPU, through Windows ML, OpenVINO, TensorRT, Qualcomm SNPE or Apple Core ML.
• Geekbench AI reports three scores, single precision, half precision and quantized, and adjusts each for accuracy against a full-precision reference, so lower output quality lowers the score.

The new Surface shows why this matters. Qualcomm's NPU rating rose from 45 to 80 TOPS, about 78% more, yet Microsoft's September 23 announcement reports up to 95% faster on-device AI, measured with Procyon AI against the first-edition Surface Pro 12-inch. TOPS got the direction right and the size wrong, as happens when memory, architecture and drivers change together.

Frequently asked questions

How can I tell whether an app is using the NPU?

On Windows 11 PCs with an NPU, Task Manager's Performance tab shows NPU utilization alongside CPU and GPU, according to Microsoft's developer guide. Run the AI feature and watch which graph rises; a flat NPU graph means the work is going elsewhere.

Why do two laptops with the same chip score differently?

Chipmakers let laptop makers configure power limits. AMD lists a configurable power range of 15–54 W for its Ryzen AI 400 chips, and Intel lists 25 W base and up to 80 W turbo power for the Core Ultra X9 388H. A thin, quiet design near the low end sustains lower clocks than a thicker one, and memory speeds differ by model too.

Does using the NPU help battery life?

It is designed to. Microsoft's developer guide says the NPU runs AI tasks more efficiently than a CPU or GPU, extending battery life, and the Windows ML documentation calls it the battery-efficient option for sustained inference. The saving applies only to apps that send their AI work to the NPU.

How do I compare a Mac with a Windows AI PC?

Use a benchmark that runs on both. Geekbench AI runs on macOS through Apple's Core ML and on Windows through ONNX and OpenVINO, and Procyon AI Computer Vision runs on Windows PCs and Macs. Compare the same score type, such as quantized against quantized.

Can a software update make an older PC Copilot+?

No. An update can bring Windows 11 to version 24H2, but the hardware minimums cannot be updated: a 40+ TOPS NPU, 16 GB of RAM and 256 GB of storage. A 2023 Intel Core Ultra 7 155H, for example, has an 11 TOPS NPU, so no update brings it to the bar.

Checks before you buy

1. Confirm the NPU figure on the chipmaker's own spec page, not an "overall" or "platform" total, and note whether it says INT8.
2. If you want Copilot+ features, confirm the minimums: 40+ NPU TOPS, 16 GB of RAM, 256 GB of storage and Windows 11 version 24H2 or newer. For Recall, also check for a Windows Hello biometric sign-in (face or fingerprint).
3. For local AI models, buy more RAM than the 16 GB minimum (32 GB leaves room for an 8-billion-parameter model alongside Windows), and check whether the laptop's memory can be upgraded later; on many thin models it cannot. Check the chip's memory bandwidth too.
4. Open the support page for each AI app you plan to use and see whether it lists your NPU, your GPU or neither.
5. Look up MLPerf Client, Procyon AI or Geekbench AI results for the exact laptop model, not just the chip.

Sources

• Microsoft: Introducing the next Surface Pro 12-inch and Surface Laptop 13-inch with Snapdragon X2 Plus (September 23, 2026)
• Qualcomm: Snapdragon X2 Plus product brief
• Qualcomm: Snapdragon X2 Elite product brief
• Qualcomm: Snapdragon X Plus product brief
• Intel: Core Ultra 7 155H specifications
• Intel: Core Ultra 7 258V specifications
• Intel: Core Ultra X9 388H specifications
• AMD: CES 2026 Ryzen AI announcements (January 5, 2026)
• Apple: Apple introduces M4 chip (May 7, 2024)
• Apple: Apple unleashes M5 (October 15, 2025)
• NVIDIA: GeForce RTX 5090 specifications
• NVIDIA: Structured sparsity in the Ampere architecture
• NVIDIA: Mastering LLM techniques, inference optimization
• Kaplan et al.: Scaling Laws for Neural Language Models
• Microsoft: Copilot+ PCs and Windows PCs, differences
• Microsoft Support: Retrace your steps with Recall
• Microsoft: Copilot+ PC feature update (April 25, 2025)
• Microsoft Learn: Copilot+ PCs developer guide
• Microsoft Learn: What is Windows ML?
• Microsoft: Windows ML is generally available (September 23, 2025)
• Ollama: Hardware support
• AMD: Ryzen AI software, LLM overview
• MLCommons: MLPerf Client benchmark
• MLCommons: MLPerf Client v2.0 (August 18, 2026)
• UL: Procyon AI Computer Vision Benchmark
• Primate Labs: Geekbench AI workloads