Skip to content

VersusCPU

NPU vs GPU for Local AI: Which Should Run Your AI Models?

NPU vs GPU for local AI: NPUs (40+ TOPS in Copilot+ PCs) are efficient for always-on Windows AI features and small models via apps such as AMD's Lemonade, while GPUs have far broader app support and their own fast VRAM, so they remain the better choice for chatbots and image generation. Microsoft is even expanding its Aion Instruct model to GPUs.

By , Editorial team

3 min read

NPU vs GPU for Local AI: Which Should Run Your AI Models?

A neural processing unit (NPU) is now a selling point on almost every new laptop and many mini PCs, and Microsoft's Copilot+ PCs have NPUs capable of 40+ trillion operations per second (TOPS). Graphics cards, meanwhile, remain the default for running local chatbots and image generators. So if you want to run AI on your own PC, which one should do the work?

This comparison is based on official documentation from Microsoft, AMD, Intel and the Lemonade project, checked on 14 November 2026. Forge PC has not benchmarked NPUs against GPUs, so we do not quote speeds. For a primer on NPUs, read what an NPU is.

The short answer

NPU

GPU (graphics card or integrated)

Designed for

Efficient, always-on AI features

Heavy parallel work, including large AI models

Power use

Low; AMD calls the NPU "more power-efficient than using the GPU"

Higher, especially discrete cards

Memory

Shares system RAM

Discrete cards have their own fast VRAM; integrated GPUs share system RAM

App support

Growing, but limited to specific apps and model formats

Broad: Ollama, LM Studio, Jan, ComfyUI and most other tools

Best for

Windows AI features, background tasks, laptops on battery

Chatbots, coding assistants, image and video generation

What NPUs are good at

NPUs are built to run AI efficiently in the background. Microsoft's Copilot+ PC page says NPUs power features such as "automatic framing in video calls" and help "refine writing in documents". Microsoft's own on-device language model is aimed at Copilot+ PCs too: in November 2026 it announced that Aion Instruct, a 2.3-billion-parameter model replacing the 3.8-billion-parameter Phi Silica, will roll out to Copilot+ PCs through the Windows Insider Program from late November. We covered that in our Aion Instruct news story.

Third-party apps can use NPUs too, but support is narrower. AMD's Lemonade server, for example, can run models on the XDNA 2 NPU in Ryzen AI 300 and 400-series processors, using specially prepared NPU and "hybrid" models rather than any model you download.

What GPUs are good at

Graphics cards remain the most flexible way to run local AI. Almost every popular app, from Ollama and LM Studio to ComfyUI, supports GPUs, and discrete cards have their own high-bandwidth video memory. As an example of raw AI throughput, Intel rates the Arc B580 graphics card at 233 TOPS (INT8), several times the 40 TOPS floor for a Copilot+ PC NPU, though TOPS figures from different vendors are not always measured the same way.

Even Microsoft is moving beyond the NPU: its Aion Instruct announcement says the model is expanding to supported GPUs, "including popular dGPUs like NVIDIA's RTX-series and AMD's RDNA4 generation".

Working together: hybrid mode

The choice is not always either/or. In Lemonade's hybrid mode on Ryzen AI PCs, AMD says the NPU processes your prompt and the integrated GPU generates the reply, playing to "each chip's strengths". Lemonade's FAQ adds that with short prompts the NPU finishes quickly, so you may see little NPU activity.

Which should you use?

  • You have a gaming PC with a graphics card: use the GPU. It will run far more apps and larger models. Our best GPU for local AI guide covers upgrades.

  • You have a Copilot+ or Ryzen AI laptop without a discrete GPU: let Windows' AI features use the NPU, and try NPU-aware apps such as Lemonade or Microsoft's Foundry Local for small models. For bigger models, the integrated GPU and plenty of RAM matter more.

  • You are buying a PC for local AI: do not pay extra for NPU TOPS alone. Memory (VRAM or unified memory) decides which models you can run, as our CPU vs GPU for local AI and RAM for local AI guides explain.

Our verdict

For running chatbots, coding assistants and image generators on your own PC in 2026, a GPU is still the better choice: it has the widest app support, and discrete cards bring their own fast memory, with Intel rating even the Arc B580 at 233 TOPS (INT8) against the 40+ TOPS NPUs in Copilot+ PCs. NPUs are best for efficient, always-on Windows AI features and small models on laptops, through apps such as AMD's Lemonade, which can also split work between NPU and GPU. Don't buy a PC for local AI on NPU TOPS alone. Forge PC has not benchmarked NPUs against GPUs; this comparison uses official documentation.

The bottom line

For running chatbots, coding assistants and image generators on your own PC in 2026, a GPU is still the better choice: it has the widest app support, and discrete cards bring their own fast memory, with Intel rating even the Arc B580 at 233 TOPS (INT8) against the 40+ TOPS NPUs in Copilot+ PCs. NPUs are best for efficient, always-on Windows AI features and small models on laptops, through apps such as AMD's Lemonade, which can also split work between NPU and GPU. Don't buy a PC for local AI on NPU TOPS alone. Forge PC has not benchmarked NPUs against GPUs; this comparison uses official documentation.

Sources

  1. 01Microsoft UK, Copilot+ PCs
  2. 02Microsoft Foundry on Windows blog, Microsoft Aion Instruct: Advancing local AI on Windows and expanding to GPU (13 November 2026)
  3. 03AMD, Getting Started with Lemonade (AMD AI Playbooks)
  4. 04Lemonade, Frequently Asked Questions
  5. 05Intel, Intel Arc B580 Graphics specifications

About the author

Forge PC Staff

Editorial team

The shared byline for articles researched, written and edited by the Forge PC editorial team. How we source, check and correct our work is set out in our editorial policy.

More from Forge PC Staff

Read next

How to Enable AMD Fluid Motion Frames (AFMF 2.1) on a Radeon Graphics Card

How-To

How to Enable AMD Fluid Motion Frames (AFMF 2.1) on a Radeon Graphics Card

To enable AMD Fluid Motion Frames, install Adrenalin 25.3.1 or newer, then select the HYPR-RX profile (RX 7000 and newer) or turn on AMD Fluid Motion Frames under Gaming > Graphics, globally or per game. AFMF 2.1 works on RX 6000 cards and newer in DirectX 11, DirectX 12, Vulkan and OpenGL games, with VSync off; RX 6000 cards need exclusive full screen.

Forge PC Staff2 min

Two vs Four RAM Sticks for DDR5: Why 2x32GB Usually Beats 4x16GB

Versus

Two vs Four RAM Sticks for DDR5: Why 2x32GB Usually Beats 4x16GB

Two DDR5 sticks are better for most builds. AMD lists DDR5-5600 as the AM5 maximum with two modules but DDR5-3600 with four, and XMP or EXPO speeds above that are overclocking and not guaranteed. Buy 2x16GB or 2x32GB for 32GB or 64GB, and use four sticks, ideally a matched kit, only if you need 128GB or more.

Forge PC Staff3 min

AMD's FSR 4.1.2 Unreal Engine 5 Plugin Brings ML Upscaling to RX 6000 Cards

News

AMD's FSR 4.1.2 Unreal Engine 5 Plugin Brings ML Upscaling to RX 6000 Cards

AMD's FSR Unreal Engine 5 plugin now uses FSR Upscaling 4.1.2, bringing ML-based upscaling to Radeon RX 6000 cards and select Ryzen integrated graphics as well as RX 7000 and RX 9000, in UE 5.3 to 5.8 games that adopt it. It's DirectX 12 only for now and adds Frame Generation 4.0.2 with driver-assisted frame pacing on RX 9000 cards.

Forge PC Staff2 min

NVIDIA's DGX Spark HPC Developer Kit Links Four Sparks Into a 512GB Desk-Side Cluster

News

NVIDIA's DGX Spark HPC Developer Kit Links Four Sparks Into a 512GB Desk-Side Cluster

NVIDIA's DGX Spark HPC Developer Kit, announced at SC26 on 16 November 2026, links four DGX Spark systems in a ring over ConnectX-7 with no network switch, for 512GB of combined memory and 20TB of NVMe storage. NVIDIA says it can spread models of up to 700B parameters across the nodes with NVFP4. No UK kit price yet; Scan sells two-Spark bundles at about £12,845.

Forge PC Staff3 min