Every major laptop chip released since 2023 — Intel’s Core Ultra, AMD’s Ryzen AI, Apple’s M-series, Qualcomm’s Snapdragon X — now includes a Neural Processing Unit alongside the CPU and GPU. Here’s what it actually does differently, and why it exists as separate silicon at all.
Why Not Just Use the CPU or GPU?
AI inference is dominated by one specific operation performed billions of times per second: matrix multiplication. CPUs are general-purpose and comparatively inefficient at this. GPUs are much better, having been built for similarly parallel graphics math — but an NPU takes it further, built from the ground up to do nothing but these specific low-precision matrix operations, at a fraction of the power draw of a GPU doing the same work.
Performance Is Measured in TOPS
TOPS (Trillion Operations Per Second) at a given precision (typically INT8) is the standard NPU benchmark. Current-generation laptop NPUs range from roughly 10-13 TOPS on earlier chips up to 40-50+ TOPS on the latest Copilot+ certified silicon.
What Actually Runs on the NPU Today
- Webcam background blur and eye contact correction during video calls
- Real-time audio noise suppression
- Live captioning and translation
- On-device image search and semantic file search (Windows Recall, Apple’s Photos search)
- Small local language models for autocomplete and quick local queries
Checking What’s Running Where
# Task Manager's Performance tab shows a dedicated NPU utilization graph # on Copilot+ certified hardware — right next to CPU, GPU, and Memory Get-CimInstance Win32_Processor | Select-Object Name, NumberOfCores
The Limitation Worth Knowing
NPUs are fast and efficient specifically for inference on already-trained, typically small-to-medium models — they’re not general-purpose compute and can’t train models from scratch the way a discrete GPU cluster can. Think of them as a highly specialized co-processor for a narrow but increasingly important class of everyday tasks, not a GPU replacement.