How Much AI Performance Do You Actually Need in a Laptop?
NPUs and AI performance numbers are everywhere. But TOPS alone do not tell you whether a laptop will handle your actual workflow. Here is what matters for everyday tasks, creator work, and running local AI models.
Image: Greg Toope editorial archive
Table of Contents
What an NPU actually does
An NPU (Neural Processing Unit) is a dedicated chip designed to run machine-learning inference tasks efficiently. Unlike a CPU, which handles general-purpose computation, or a GPU, which excels at parallel graphics rendering, an NPU is optimized specifically for the matrix math that neural networks require. The benefit is not raw speed but efficiency: an NPU can handle AI workloads at a fraction of the power draw of a GPU doing the same work.
In practical terms, your NPU handles tasks like real-time background blur in video calls, live captioning and transcription, camera effects (eye contact correction, portrait mode), voice noise suppression, and on-device text prediction. These are the features you use without thinking about them. The NPU runs them in the background while your CPU and GPU handle everything else.
What an NPU does not do is make every AI app faster. If you are using a cloud-based AI tool like ChatGPT or Google Gemini, your laptop's NPU is irrelevant — the processing happens on remote servers. The NPU only matters for local AI workloads, where the model runs on your machine. This distinction is critical because most laptop marketing conflates "AI performance" with cloud AI, which has nothing to do with your hardware.
TOPS: what the number means (and does not)
TOPS (Tera Operations Per Second) is the headline metric for NPU performance. It measures how many trillion operations the NPU can process per second. Microsoft's Copilot+ PC program requires a minimum of 40 TOPS from the NPU for certification. Here is where the major platforms stand as of 2026:
| Platform | NPU TOPS | Notes |
|---|---|---|
| Apple M5 (base) | ~38 TOPS (Neural Engine) | 16-core Neural Engine; Apple does not publish TOPS directly |
| Apple M5 Pro / M5 Max | Higher (undisclosed) | Enhanced Neural Accelerators in GPU cores; 307 GB/s memory bandwidth on Max |
| Intel Core Ultra Series 2 (Lunar Lake) | 48 TOPS | Copilot+ certified; 4P + 4E cores |
| Intel Core Ultra Series 3 (Panther Lake) | 50 TOPS | CES 2026 debut; up to 16 CPU cores |
| AMD Ryzen AI 300 | 50 TOPS | XDNA 2 architecture; up to 128 GB RAM support on PRO |
| AMD Ryzen AI 400 | 60 TOPS | CES 2026; Zen 5 cores, RDNA 3.5 graphics |
The problem with TOPS is that it measures peak throughput on a narrow benchmark, not real-world application performance. A 60-TOPS NPU sounds impressive, but if your software does not support it, or if your bottleneck is RAM (which it usually is), that number is meaningless. Think of TOPS like engine horsepower: it tells you something about capability, but not whether the car is good for your commute.
Three workload levels: which one are you?
Level 1 — Everyday assistance (minimal AI hardware needed). If your AI use is background blur on Zoom, live captions, voice noise suppression, and maybe Copilot text suggestions in Windows or Apple Intelligence summaries, any modern NPU will do. You do not need 50+ TOPS. A laptop with a 40-TOPS NPU and 16 GB of RAM will handle all of this without breaking a sweat. Battery life matters more than NPU speed here because these tasks run continuously in the background.
Level 2 — Creator acceleration (GPU and RAM matter more than NPU). If you edit video, work with photos, or use AI-assisted creative tools (Photoshop Generative Fill, DaVinci Resolve Magic Mask, Final Cut Pro subject isolation), the NPU helps but is not your primary engine. Your GPU and RAM matter far more. Video editing with AI features like auto-reframing, object tracking, and noise reduction benefits from a strong GPU and at least 32 GB of unified memory. A Mac Studio M4 Max with 32 GB handles 4K ProRes timelines with AI effects without struggling. The NPU assists with background tasks, but the GPU does the heavy lifting.
Level 3 — Local LLMs (RAM is king, NPU is secondary). If you want to run large language models locally — Llama 3.1, Gemma 4, DeepSeek, Mistral — your NPU matters less than your RAM capacity. Local LLMs must fit entirely in memory to run at usable speeds. A 7B parameter model (like Llama 3.1 8B) needs roughly 8 GB of dedicated memory at 4-bit quantization. A 13B-14B model needs 12-16 GB. A 70B model needs 35-48 GB. If your laptop has 8 GB of RAM, no amount of NPU TOPS will let you run a useful model. The NPU accelerates inference once the model fits, but memory capacity is the gatekeeper.
RAM: the real bottleneck for local AI
If you care about running AI locally, stop looking at TOPS and start looking at RAM. Here is a practical guide based on tested configurations:
| RAM | What you can run | Realistic use case |
|---|---|---|
| 8 GB | Small models (3B or less, quantized) | Barely usable for local LLMs; fine for cloud AI + everyday NPU tasks |
| 16 GB | 7B models (Llama 3.1 8B, Gemma 9B at 4-bit) | Entry point for local AI; model runs but leaves little room for other apps |
| 32 GB | 13B-14B models, multiple smaller models simultaneously | Recommended for serious local AI use; model + apps coexist comfortably |
| 64 GB+ | 30B-70B models (quantized), fine-tuning | Power users; can run capable models alongside creative workloads |
Apple Silicon has a unique advantage here: unified memory architecture means the GPU can access nearly all system RAM. A Mac Studio M4 Max with 64 GB of unified memory can allocate a large chunk of that to a local model. On Windows laptops with discrete GPUs, you are limited by VRAM (typically 8-16 GB on consumer GPUs). AMD's Ryzen AI Max+ chips with support for up to 96 GB of unified system memory address this gap, but they are still rare in consumer laptops.
Apple vs Intel vs AMD: where each shines
Apple Silicon (M4/M5) wins on efficiency and unified memory. If you want to run local LLMs with a large model, a Mac with 64 GB+ of unified memory is the most practical option because the GPU can address all of it. Apple's Neural Engine is capable but less transparent about performance metrics. Apple Intelligence features (summarization, writing tools, image generation) are tightly integrated and run locally on the Neural Engine.
Intel (Core Ultra Series 2/3) offers the broadest software compatibility and Copilot+ certification. Lunar Lake (Series 2) at 48 TOPS meets the Copilot+ threshold. Panther Lake (Series 3) pushes to 50 TOPS with significant multi-threaded performance improvements. Intel's advantage is Windows ecosystem compatibility and broad OEM availability — you can find Intel AI laptops at every price point.
AMD (Ryzen AI 300/400) offers the highest raw NPU TOPS (50-60) and strong multi-core CPU performance. The Ryzen AI 300 PRO supports up to 128 GB of RAM, which is exceptional for local AI workloads. AMD's XDNA 2 NPU architecture is efficient, and the integrated Radeon 890M graphics are competitive. The weakness is OEM availability — fewer premium laptop models use AMD AI chips compared to Intel.
A practical buying rule
For a creator carrying action-camera footage and editing daily — like I do with my Mac Studio M4 Max — a balanced laptop with enough memory and fast storage is usually better than one with an impressive AI number and a cramped configuration. Here is my decision framework:
If you only use cloud AI (ChatGPT, Gemini, Copilot): Any modern laptop with 16 GB RAM is fine. Ignore NPU specs entirely. Spend on screen quality, battery life, and storage instead.
If you use AI-assisted creative tools (Photoshop, Resolve, Final Cut): Prioritize GPU performance and 32 GB+ of RAM. The NPU helps with background tasks but is not your bottleneck. A Mac Studio M4 Max or a Windows laptop with a discrete GPU and 32 GB RAM will serve you well.
If you want to run local LLMs: Maximize RAM first. Get 32 GB minimum, 64 GB if budget allows. NPU TOPS is secondary. On Windows, look for AMD Ryzen AI Max+ with high unified memory support. On Mac, any M-series with 64 GB unified memory is excellent for this use case.
If you want all of the above: An Apple Silicon Mac with 64 GB unified memory remains the most versatile option as of 2026. It handles creative workloads, runs local LLMs using unified memory, and delivers excellent battery life. The trade-off is price and limited upgradeability.
Buy for the app you use weekly, not the benchmark number on a spec sheet.
Related Articles
How Much Storage Does Video Editing Really Need in 2026? Do You Need a Motorcycle Airbag Vest? The Surprising Data Is Local AI Worth Running on a Laptop in 2026? Laptop Guide Best Laptop Extender Displays for Remote Work 2026FAQ
Not everyone does. You need one when supported AI features solve a real recurring task — whether that is local transcription, on-device LLMs, or AI-assisted creative work. If your AI use is entirely cloud-based, any modern laptop works.
TOPS measures raw NPU throughput. Microsoft requires 40 TOPS for Copilot+ certification. Current chips range from 40-60 TOPS. Higher TOPS helps sustained AI workloads but does not matter for casual use like background blur or live captions.
For most workflows, no. A high-TOPS NPU with only 8 GB of RAM cannot run meaningful local models, while a modest NPU with 32 GB of RAM can run capable models. RAM capacity is the primary bottleneck for local AI.
No. ChatGPT runs on OpenAI's cloud servers. Your NPU does not help with cloud AI services. To run AI locally, you need open-weight models like Llama 3.1, Gemma, or Mistral, plus sufficient RAM to hold them.
Yes. Apple calls it the Neural Engine. The M5 has a 16-core Neural Engine integrated into the SoC. Apple does not publish TOPS numbers directly, but independent estimates put base M5 at approximately 38 TOPS, with Pro and Max variants higher.
Information checked August 27, 2026. Confirm current specifications, availability, and pricing with the manufacturer or an authorized Canadian dealer.