Every laptop spec sheet in 2026 now includes a TOPS figure for the NPU. Forty TOPS. Fifty TOPS. Some chips are pushing past 100. The number keeps going up, the marketing keeps getting louder — and most buyers have no real idea what they’re comparing. That’s not their fault. The way the industry communicates this metric is, frankly, a mess.
I’ve spent the better part of two years testing laptops with dedicated NPUs — from Qualcomm’s first Snapdragon X Elite machines to the latest Intel Core Ultra 200H systems — and the gap between what manufacturers claim and what users actually experience is significant enough to warrant a proper technical breakdown. This is that breakdown.
What “NPU” Actually Means
NPU stands for Neural Processing Unit. It’s a dedicated processor core designed specifically to accelerate matrix multiplication operations — the mathematical foundation of virtually every modern AI model. Unlike a CPU, which handles general-purpose sequential tasks, or a GPU, which parallelizes thousands of simpler floating-point operations, an NPU is purpose-built to execute the specific tensor operations that neural networks require, using far less power than either alternative.
The key word there is dedicated. Your CPU can run neural network inference. Your GPU can too, and much faster for large models. But both consume significant power to do it. An NPU handles narrow, well-defined AI tasks at a fraction of the wattage — which matters enormously in a laptop where battery life is a real constraint.
An NPU doesn’t replace your GPU for AI workloads. It handles specific, recurring, low-complexity inference tasks so that your GPU and CPU don’t have to — preserving performance headroom for everything else you’re running simultaneously.
What TOPS Measures — and What It Doesn’t
TOPS stands for Tera Operations Per Second. One TOPS equals one trillion operations per second. It’s a throughput metric — it tells you how many mathematical operations the NPU can execute in a given second under ideal conditions.
The problem is “ideal conditions.” TOPS figures are almost always measured at peak theoretical throughput using INT8 precision (8-bit integer arithmetic), which is the format that maximizes raw operation count. Real-world AI workloads often use FP16 or FP32 precision, which reduces effective throughput significantly — sometimes by half or more depending on the chip’s architecture.
This means you cannot directly compare TOPS across different chip architectures and assume the higher number wins. A 45 TOPS NPU from one vendor and a 45 TOPS NPU from another may perform very differently on the same real-world task, because the underlying architecture, memory bandwidth, and supported precision formats differ. The number is a starting point for comparison, not a verdict.
The 40 TOPS Threshold: Why It Exists and What It Unlocks
Microsoft drew a line in the sand with Copilot+ PC certification: 40 TOPS NPU performance, minimum. Below that threshold, a device doesn’t qualify for the Copilot+ feature set — which includes Windows Studio Effects (live background blur, eye contact correction, automatic framing during video calls), Cocreator in Paint, live captions with real-time translation, and the Recall feature that creates a searchable timeline of your activity.
That 40 TOPS number isn’t arbitrary. Microsoft’s engineers benchmarked the compute requirements of these features running concurrently with a typical workload — browser, communication apps, video call — and determined that 40 TOPS is the floor at which the NPU can handle AI tasks without measurably impacting overall system performance. Below that, you’re offloading AI tasks to the CPU, which raises power consumption and can cause perceptible slowdowns.
What this means practically: if you’re buying a Windows laptop in 2026 and you care about the AI-assisted features Microsoft is building into the OS, a Copilot+ certified chip is a hard requirement, not a nice-to-have. The features literally don’t run on non-certified hardware.
What Tasks the NPU Actually Handles Day-to-Day
This is where most explanations go wrong. They list “AI tasks” without being specific about which ones run on the NPU versus the GPU versus the CPU. Let me be precise.
Tasks that run on the NPU by design — meaning the operating system or application explicitly routes them there — include: real-time background removal in video calls, noise suppression for microphone audio, eye contact correction, automatic framing, on-device speech recognition for live captions, and Windows Hello facial authentication. These are all lightweight, continuous inference tasks that run in the background constantly whenever you’re in a video call or using voice features.
Tasks that do NOT meaningfully benefit from the NPU include: running local LLMs like Llama 3 or Mistral (these are GPU-bound for any model above 3B parameters), Stable Diffusion image generation (GPU-bound entirely), video encoding and transcoding (handled by dedicated media encoders, not the NPU), and 3D rendering of any kind.
If you’re buying a laptop specifically to run local AI models — Ollama, LM Studio, llama.cpp — the NPU TOPS figure is almost irrelevant to your decision. What matters is GPU VRAM and unified memory bandwidth. A MacBook Pro M4 with 24GB unified memory will run local LLMs circles around a Copilot+ PC with a faster NPU but only 16GB of shared VRAM.
NPU vs GPU: Which Runs What, and Why It Matters for Battery
The reason dedicated NPUs exist — rather than just routing everything to the GPU — comes down to power efficiency. Running a continuous background task like noise suppression on a discrete GPU would consume somewhere between 10W and 30W depending on the GPU and the precision used. The NPU handles the same task at under 1W in most implementations. Over an eight-hour workday with video calls, that difference compounds into hours of battery life.
Think about what happens if you’re on a three-hour video call while also running a browser with fifteen tabs and a Slack workspace. On a laptop without a dedicated NPU, background blur and noise suppression are running on the CPU — taking CPU cycles away from everything else and generating heat. On a Copilot+ device, those tasks are silently offloaded to the NPU, and your CPU remains free for the applications that need it. The experience difference is real and measurable.
Where the GPU still dominates is in burst AI inference — anything that isn’t continuous and lightweight. Running a local AI coding assistant that generates completions on demand, processing a batch of images through a vision model, or using an app like DaVinci Resolve’s AI noise reduction on video footage — all of these are GPU workloads. The NPU isn’t fast enough for burst tasks that require the full compute of a modern GPU architecture.
How to Evaluate NPU Performance Beyond the TOPS Number
Since raw TOPS can be misleading, here’s what to actually look for when comparing NPU implementations across laptops:
Supported precision formats. Check whether the NPU supports INT4 and INT8 inference natively. Many newer models — especially quantized LLMs — run in INT4 or INT8 to reduce memory footprint. If the NPU only supports FP16, it can’t run these efficiently. Qualcomm’s Hexagon NPU in the Snapdragon X series has strong INT4 support; Intel’s NPU in Core Ultra 200 series has improved INT8 support but is weaker on INT4.
Memory bandwidth. An NPU starved of memory bandwidth can’t hit its theoretical TOPS in practice. Unified memory architectures like Apple Silicon have a significant advantage here — the NPU, CPU, and GPU share a high-bandwidth memory pool. On most Intel and AMD platforms, the NPU pulls from system RAM at standard LPDDR5X speeds, which can become a bottleneck on larger models.
Software ecosystem. The NPU is only useful if applications actually target it. Qualcomm has the most mature developer ecosystem for Windows NPU workloads via the AI Hub platform. Intel’s OpenVINO toolkit provides decent coverage. AMD’s Ryzen AI software stack is the youngest and still catching up. Apple’s Core ML is the most mature ecosystem of all but only relevant within macOS.
Should NPU TOPS Drive Your Laptop Buying Decision in 2026?
If you’re primarily a Windows user who relies on video calling, live transcription, and OS-level AI features: yes, meeting or exceeding the 40 TOPS Copilot+ threshold is worth prioritizing. The real-world quality difference in video call features on a Copilot+ device versus a non-certified one is immediately noticeable.
If you’re buying for local AI model inference, machine learning development, or GPU-accelerated creative work: NPU TOPS is not the metric that should drive your decision. GPU VRAM capacity, unified memory architecture, and CPU core count matter far more. A laptop with a 45 TOPS NPU but only 8GB of shared GPU memory is a worse local AI machine than one with a 30 TOPS NPU and 16GB dedicated VRAM.
If you’re buying a MacBook: the TOPS comparison with Windows machines is largely academic. Apple’s Core ML framework is more mature than any Windows NPU software stack, and the unified memory architecture means Apple Silicon can run larger models in the NPU+CPU+GPU pipeline more efficiently than the raw TOPS number implies. The M4’s 38 TOPS frequently outperforms higher-TOPS Windows chips on the specific tasks Core ML targets.
The bottom line: use TOPS as a threshold check — make sure you’re above 40 if you want Copilot+ features — and then look at memory, software ecosystem, and real-world benchmark results for the specific workloads you actually run. The number is a floor, not a ceiling, and certainly not a complete performance story.
For a practical comparison of current AI laptops ranked by real-world NPU and overall AI performance, see our Best AI Laptops 2026 roundup, updated monthly.
Frequently Asked Questions
Does a higher NPU TOPS always mean better AI performance?
No. TOPS is measured under ideal conditions at a specific precision format (usually INT8). Real-world performance depends on the precision formats supported, memory bandwidth available to the NPU, and whether the applications you use actually target that NPU via platform-specific frameworks. A 50 TOPS NPU on a poorly-optimized software stack can underperform a 40 TOPS NPU with mature driver support.
Can the NPU run local LLMs like Llama 3?
On most current Windows laptops, the NPU is too limited for LLMs above about 1–3 billion parameters at acceptable speeds. Qualcomm’s Snapdragon X series is the exception — its Hexagon NPU can run quantized 7B models at usable speeds via Qualcomm’s AI Hub SDK. For everything else, local LLM inference remains GPU-bound. Apple Silicon is a special case where the NPU contributes to inference via Core ML’s heterogeneous compute pipeline.
What is the Copilot+ PC requirement exactly?
Microsoft requires a minimum of 40 TOPS dedicated NPU performance, 16GB of RAM, and 256GB of storage for Copilot+ certification. The 40 TOPS threshold is the critical one — it’s the compute floor Microsoft determined is necessary to run AI features concurrently without degrading overall system performance.
Is the NPU the same as the GPU on integrated graphics?
No. They are separate silicon blocks on the same chip die. The integrated GPU handles graphics rendering and GPU-compute tasks. The NPU handles dedicated neural network inference. They operate independently and can run simultaneously — which is exactly the point. Background AI tasks on the NPU leave the GPU free for graphics and heavier compute workloads.
Should I wait for higher NPU TOPS before buying?
If you need a laptop now and the device you’re considering meets the 40 TOPS Copilot+ threshold, the current NPU tier is sufficient for all currently-shipping OS-level AI features. NPU performance will continue increasing, but the software catching up to even current hardware is the slower constraint. Waiting for more TOPS rarely makes sense unless you have a specific future workload that requires it.
WRITTEN BY
Alex Carter
Senior Tech Editor — AI GPUs & Workstations
8 years covering AI hardware and GPU architecture. Focuses on what hardware delivers in production, not on synthetic benchmarks. Has tested laptops from every major Copilot+ platform since the first Snapdragon X Elite devices shipped.
Specialties: NVIDIA & AMD GPUs · AI inference benchmarking · Workstation builds · Local LLM deployment
Stay up to date with the latest AI hardware reviews, buying guides, and analysis at AiGigabit.com. Bookmark us for daily updates.
