In this article
- What an NPU does
- What a GPU does for AI
- How the numbers compare
- Efficiency against throughput
- Memory decides model size
- Software support and the Copilot+ line
- Which block runs which task
- Common mistakes when reading AI specs
- Questions to ask before you buy
- Desktop view
- What this means when buying
- Frequently asked questions
- Sources & credits
Short answer
An NPU is a small, power-efficient AI block inside a chip. A GPU is a large parallel processor that runs bigger models faster. Intel lists the Core Ultra 7 258V with 47 NPU TOPS and 64 GPU TOPS, so even on an AI-focused laptop the graphics block supplies more raw compute. Use the NPU for light, always-on tasks and the GPU for heavy work.
- NPUs are built for low power; GPUs are built for throughput.
- Where makers list both, GPU TOPS are often larger than NPU TOPS.
- Microsoft sets 40 TOPS as the NPU bar for many Copilot+ features.
- For local language models, memory size usually decides what you can run.
What an NPU does#
A neural processing unit is a block designed to run the repeated multiply-and-add steps that neural networks use. It does this with far less power than a CPU or a GPU would for the same job, and that is its purpose. In a laptop, it keeps a background feature such as voice isolation, camera effects or live captions running without draining the battery or spinning up fans.
An NPU is fixed in size and tied to the chip. Intel's Core Ultra 9 285K lists 13 TOPS of NPU, the Ryzen AI 9 HX 370 lists 50, and Qualcomm's Snapdragon X Elite lists 45. You cannot add one to a desktop, and you cannot upgrade it later. A buyer who wants more AI compute on a desktop adds a graphics card, not an NPU. What you get is decided on the day you buy the chip.
Apple takes a similar approach with its Neural Engine, which Apple lists at 16 cores on the M4 and M5, and which the company rates at 38 trillion operations per second on the M4. Apple gives no TOPS figure for the M5's Neural Engine, and it says the M5's GPU carries more of the AI compute than before.
Software support is the catch. A program must be written or converted to use the NPU through a framework such as OpenVINO or ONNX Runtime, and not every model or operation is supported. When a model does not map cleanly, the work falls back to the CPU or the GPU.
What a GPU does for AI#
A GPU has thousands of small units that run in parallel, which suits neural networks. It also has fast memory of its own, called VRAM on a graphics card, and that memory is the main limit on model size. Large language models, image generators and training jobs run on GPUs because they need both the compute and the memory bandwidth.
Graphics chips come in two forms. An integrated GPU shares system memory and sits inside the CPU package, as on most laptops. A discrete GPU is a separate card with its own VRAM. Intel lists an integrated GPU at 64 TOPS INT8 on the Core Ultra 7 258V, 75 TOPS on the 265H and 122 TOPS on the Core Ultra X9 388H. A discrete card goes well beyond that. For sizing a card, read our VRAM guide for local LLMs and the best GPU for local LLM guide.
How the numbers compare#
Makers list TOPS for each block, which lets you see the split on a single chip. The table below uses figures from each maker's own page.
| Chip | NPU TOPS | GPU TOPS | Platform TOPS |
|---|---|---|---|
| Intel Core Ultra 7 258V | 47 | 64 | 115 |
| Intel Core Ultra 7 265H | 13 | 75 | 97 |
| Intel Core Ultra X9 388H | 50 | 122 | Not stated on the SKU page |
| Intel Core Ultra 9 285K (desktop) | 13 | 8 | 36 |
| AMD Ryzen AI 9 HX 370 | 50 | Not stated | 80 |
All Intel figures are INT8 and come from Intel's specification pages for the 258V (opens in a new tab), the 265H (opens in a new tab), the 388H (opens in a new tab) and the 285K (opens in a new tab). AMD lists the HX 370 (opens in a new tab) at 50 NPU TOPS and 80 TOPS overall, without a separate GPU figure.
The pattern is plain. On the 258V the NPU supplies less than half of the platform total. On the 265H it supplies about one-seventh. Only on the 285K desktop chip does the NPU out-rate the integrated GPU, and there both are small, because a desktop build is meant to have a discrete graphics card.
Efficiency against throughput#
Peak TOPS describe speed, not efficiency. The reason to have an NPU at all is that it does a small job on a small power budget. A GPU can run the same job faster, but it uses more power and wakes the larger part of the chip. For a call that runs for an hour with background noise removal, the NPU is the better tool. For a model that generates images or text on demand, the GPU is.
Independent data on this trade-off is limited. ComputerBase ran UL Procyon AI Computer Vision on an older Intel Core Ultra 7 165H and recorded 358 points on the NPU, 600 on the integrated GPU and 233 on the CPU. That puts the GPU well ahead on raw speed for that model. A Procyon score measures speed only, so it says nothing about how much power each block used.
Intel's own run of MLPerf Client on a Core Ultra 9 288V NPU gave 1.09 seconds to first token and 18.55 tokens/s on Llama 2 7B. That is a vendor result, and it shows that an NPU can run a seven-billion-parameter model at a readable speed. It does not say how a GPU would compare, since Intel gave no numeric GPU result.
Memory decides model size#
Compute is rarely the first limit for local language models. Memory is. A model must fit in the memory the processor can reach, and the speed of that memory sets how quickly text is generated. A discrete graphics card has its own VRAM, while an integrated GPU or an NPU uses system memory shared with the CPU.
The memory limits differ by chip. Intel lists up to 32 GB of LPDDR5X-8533 for the 258V and up to 128 GB for the 265H. AMD lists up to 128 GB of LPDDR5x-8000 on the Ryzen AI Max+ 395, which has a 256-bit memory bus. Apple lists up to 32 GB of unified memory at 153 GB/s for the M5. A laptop with a large pool of fast shared memory can run models that a small discrete card cannot, though more slowly than a high-end card.
Software support and the Copilot+ line#
In its developer guide, Microsoft puts the NPU line for many new Windows AI features at 40 TOPS. That is why NPU size matters for a Copilot+ PC and matters little for developers who run PyTorch or llama.cpp, which target the CPU or GPU. Intel's 13 TOPS NPUs fall below the line, and the 47 TOPS 258V and 50 TOPS HX 370 clear it.
On the GPU side, support is broad. CUDA on NVIDIA cards is the most widely supported path, and AMD's ROCm is a growing alternative. Our CUDA vs ROCm explainer covers the difference. On the NPU side, support depends on the chip maker's toolkit, so check that your software lists your NPU.
Which block runs which task#
In everyday use the split looks like this. Background features that run for long stretches suit the NPU: removing noise from a call, blurring or framing a camera feed, and generating live captions. These tasks use small models, run constantly and gain the most from low power.
Interactive tasks that need quick, bursty speed suit the GPU. These include chatting with a local language model, generating an image, upscaling video and running a code assistant. They use larger models and need memory bandwidth more than they need a low power draw.
Training and fine-tuning belong to the GPU, and usually to a discrete card with plenty of VRAM. No NPU in the chips above is designed for it. If your plan is to train or fine-tune, read the deep learning GPU picks before you look at any NPU figure.
The CPU still has a role. It runs the program, prepares data and handles any model operation that the NPU or GPU does not support. A weak CPU can slow a good GPU, but a strong CPU does not replace one.
Common mistakes when reading AI specs#
- Treating NPU TOPS as total AI speed. Intel lists the 258V at 47 NPU TOPS and 115 platform TOPS, so the NPU is only part of the picture.
- Comparing TOPS across makers without checking precision. Intel labels its figures INT8, and the others do not state one beside the NPU number.
- Assuming a bigger NPU means faster chat. A local language model is limited by memory size and bandwidth first.
- Expecting to add an NPU later. It is part of the chip and cannot be upgraded.
- Ignoring software. An NPU helps only when the program you use supports it.
Questions to ask before you buy#
Start with the use. If the answer is video calls, document search and light assistants, a chip whose NPU clears Microsoft's bar is a sensible target. If it is running models of your own, ask how much memory the machine can hold and how fast that memory is. If it is training or fine-tuning, ask how much VRAM a discrete card offers.
Then ask about software. Check whether the tools you plan to use list support for the NPU, the integrated GPU or only CUDA. Support for CUDA, which runs on NVIDIA cards, is broader than support for any NPU at present. AMD's ROCm and Intel's OpenVINO are options on their own hardware.
Last, ask about the product, not the family. Chip families contain parts with very different NPUs, and laptop makers sell several memory sizes under one model name. Check the exact chip and the exact configuration on the maker's page.
Desktop view#
For a desktop build, the question is simpler. Desktop chips seldom carry a large NPU, and the graphics card does the AI work. Phoronix put AMD's Ryzen 9000 chips ahead of Intel's Arrow Lake in CPU AI tests, and Puget's MLPerf Client run with a GPU saw the CPU change little, which tells you where the effort belongs. The AI and ML CPU guide covers the chip side.
What this means when buying#
If you want a laptop for Copilot+ features, check the NPU figure on the exact chip and see whether it reaches Microsoft's bar. If you want to run local models, look first at memory size and the GPU, then treat the NPU as an extra. Our AI PC processors compared article lists the NPU and platform TOPS of the main laptop chips, and the best Copilot+ PC guide lists machines that meet Microsoft's line.
If you are building a desktop, put the money into the GPU. The deep learning GPU picks cover that choice, and the AI processors page lists every CPU guide on this site.
If you buy an AI laptop in India for Copilot+ features, confirm the NPU rating on the maker's Indian product page, because retail listings sometimes mix up different regional configurations.
Frequently asked questions#
Is an NPU better than a GPU for AI?
Not for speed. An NPU is built for low power on small, steady tasks. A GPU runs larger models faster. The NPU wins on battery life for background tasks, while the GPU wins for heavy model work.
Can an NPU run large language models?
It can run smaller ones. Intel reported 18.55 tokens/s on Llama 2 7B for a Core Ultra 9 288V NPU in its own MLPerf Client run.
Do I need an NPU in a desktop?
No. Desktop AI work runs on the GPU and CPU. Intel lists 13 TOPS on the Core Ultra 9 285K and AMD's page shows no NPU on Ryzen 9000 desktop chips.
What does a Copilot+ PC need?
Per Microsoft's developer guide, 40 TOPS is the NPU level that many new Windows AI features call for. A PC below that level can still run Windows, but those features may not be offered on it.
Does more NPU TOPS mean faster local LLMs?
Not by itself. Memory size, memory speed and software support also decide how a model runs, and most local LLM tools use the CPU or GPU, not the NPU. TOPS is a poor guide to language-model speed.
Can I add an NPU later?
No. An NPU is part of the processor chip, so it is fixed when you buy. If you want more AI performance later, you can add a graphics card to a desktop, but a laptop cannot be changed this way.
Sources & credits#
We have not tested these products ourselves. Specs come from the makers; benchmark figures come from the sources we name.
- Intel, Core Ultra 7 258V specifications (opens in a new tab)
- Intel, Core Ultra 7 265H specifications (opens in a new tab)
- Intel, Core Ultra X9 388H specifications (opens in a new tab)
- Intel, Core Ultra 9 285K specifications (opens in a new tab)
- AMD, Ryzen AI 9 HX 370 product page (opens in a new tab)
- Microsoft Learn, Copilot+ PCs developer guide (opens in a new tab)
- ComputerBase, AI benchmarks for Snapdragon X Elite, Ryzen 8000 and Core Ultra (opens in a new tab)
- Intel Newsroom, MLPerf Client NPU support (opens in a new tab)
- Puget Systems, Intel Core Ultra 200S Plus content creation review (opens in a new tab)
- Phoronix, AMD Ryzen 9000 vs. Intel Core Ultra Arrow Lake On Linux (opens in a new tab)
- Apple Newsroom, M5 chip (opens in a new tab)
- AMD, Ryzen AI Max+ 395 product page (opens in a new tab)
Facts and figures in this article come from these sources. Articles are compiled from the makers' published specifications and the benchmarks named above, and checked by the NextAIGrid Editorial Team.
Disclaimer
This explainer is for information only. Specifications change, so check the maker page. Found an error? Tell us.
How was this explainer put together?
It draws on maker specification pages and published third-party benchmarks, named in the article. We have not tested these products ourselves. Specs come from the makers; benchmark figures come from the sources we name. Specifications and sources were last checked on 17 November 2026. Read how we research and our editorial policy.
Spotted an error? Tell us. Confirmed fixes are listed on the corrections page.