Explainer

AI PC Processors Compared: NPUs and TOPS

In this article
  1. What an AI PC processor actually contains
  2. NPU and platform TOPS across the main chips
  3. The chip families in plain terms
  4. How to read a maker's AI claim
  5. Why TOPS can mislead
  6. What published benchmarks show
  7. The Copilot+ line and what it means
  8. Desktop processors and the NPU
  9. What this means when buying
  10. Frequently asked questions
  11. Sources & credits

Short answer

The NPU figure on a laptop chip tells you less than most buyers assume. AMD lists 50 TOPS for the Ryzen AI 9 HX 370, Intel lists 47 for the Core Ultra 7 258V and Qualcomm lists 45 for the Snapdragon X Elite. Microsoft asks for 40 or more for Copilot+ features. In ComputerBase's Procyon run, the graphics chip beat the NPU on an older Intel part.

  • Intel labels its TOPS as INT8; AMD, Qualcomm and Apple do not state the precision.
  • Platform TOPS add CPU, GPU and NPU, and the GPU often supplies more than the NPU.
  • Desktop chips differ: 13 TOPS on the Core Ultra 9 285K, no NPU on Ryzen 9000.
  • Choose by the software you run, not the biggest TOPS figure.

What an AI PC processor actually contains#

An AI PC processor is a normal chip with three places to run AI code. The CPU runs everything, and handles small models well. The GPU has many parallel units and handles larger, heavier models. The NPU is a block built for the repeated matrix maths in neural networks, designed to do it at low power.

The point of an NPU is efficiency. A laptop can hold a background task, such as noise removal on a call or live captions, on the NPU while the battery lasts longer and the fans stay quiet. It is not designed to beat a graphics chip on large models. Our NPU vs GPU explainer covers that split in more detail.

Makers quote TOPS, trillions of operations per second, as the headline figure. It is a peak number under ideal conditions for a stated precision. Real software rarely reaches it, and the same chip can score differently under different frameworks. That caveat applies to every row below.

NPU and platform TOPS across the main chips#

The table collects figures from each maker's own page. "Platform" is the total the maker quotes for CPU, GPU and NPU together, where it gives one.

Table
ChipNPU TOPSPlatform TOPSPrecision stated
AMD Ryzen AI 9 HX 3705080Not stated by AMD
AMD Ryzen AI Max+ 39550126Not stated by AMD
AMD Ryzen AI 9 HX 4756091Not stated by AMD
Intel Core Ultra 7 258V47115INT8
Intel Core Ultra 7 265H1397INT8
Intel Core Ultra X9 388H50Not stated on the SKU pageINT8
Qualcomm Snapdragon X Elite45Not statedNot stated by Qualcomm
Qualcomm Snapdragon X Plus45Not statedNot stated by Qualcomm
Qualcomm Snapdragon X2 EliteUp to 85Not statedNot stated by Qualcomm
Apple M438Not applicableNot stated by Apple
Apple M5Not stated by AppleNot applicableNot stated by Apple

Maker sources: AMD lists the Ryzen AI 9 HX 370 (opens in a new tab), the Ryzen AI Max+ 395 (opens in a new tab) and the Ryzen AI 9 HX 475 (opens in a new tab). Intel lists the Core Ultra 7 258V (opens in a new tab), the Core Ultra 7 265H (opens in a new tab) and the Core Ultra X9 388H (opens in a new tab). Qualcomm lists the Snapdragon X Elite (opens in a new tab), the X Plus (opens in a new tab) and the X2 Elite (opens in a new tab). Apple's figures come from its M4 (opens in a new tab) and M5 (opens in a new tab) announcements.

Three things stand out. First, the 265H lists only 13 TOPS of NPU even though Intel puts it in the same Core Ultra 200 family as the 47 TOPS 258V. Second, the platform totals are far bigger than the NPU totals: Intel lists 97 TOPS for the 265H and 115 for the 258V, and AMD lists 80 for the HX 370. Third, Apple gives a figure for the M4's Neural Engine, but its M5 announcement says only that the Neural Engine is improved and that GPU AI compute is more than four times the M4's.

The chip families in plain terms#

AMD's Ryzen AI 9 HX 370 has 12 cores and 24 threads, made of four Zen 5 cores and eight smaller Zen 5c cores, with a Radeon 890M graphics block of 16 graphics cores. AMD lists a 28 W default power with a configurable range of 15 to 54 W. The Ryzen AI Max+ 395 is the large version: 16 cores, a Radeon 8060S with 40 graphics cores, and 256-bit LPDDR5x-8000 memory of up to 128 GB. That memory bandwidth and capacity is what makes it interesting for local language models.

Intel's Core Ultra 7 258V has eight cores in a 4P plus 4 low-power layout and an Arc 140V GPU rated at 64 TOPS INT8. Its memory is LPDDR5X-8533 with a 32 GB maximum, which limits model size. The 265H is a larger H-series chip with 16 cores, an Arc 140T GPU at 75 TOPS and memory up to 128 GB. The 388H adds a GPU rated at 122 TOPS INT8.

Qualcomm's Snapdragon X Elite uses 12 Oryon cores with an Adreno GPU, and Apple's M4 has up to ten cores with a 16-core Neural Engine. Apple's M5 lists unified memory bandwidth of 153 GB/s and up to 32 GB. These details matter because a local model must fit in memory, and speed depends on memory bandwidth as much as on TOPS.

How to read a maker's AI claim#

Start with the exact chip, not the family. The same brand name covers parts with very different NPUs, as the Core Ultra 200 family shows. Then check whether the figure is NPU only or the whole platform. Next, look for the precision, and treat a missing label as a reason for caution. Finally, find independent results with the software you intend to use.

A short checklist helps:

  • Is the NPU figure at least 40 TOPS if you need Copilot+ features?
  • Does the chip have enough memory for the models you plan to run?
  • Is there an independent benchmark for the software you use?
  • Does the maker list the figure on its own product page?

Why TOPS can mislead#

A TOPS figure depends on precision. The same hardware usually shows a larger number at a lower precision, so a figure quoted at INT8 can look bigger than one quoted at a higher precision. Intel states INT8 on every figure. AMD, Qualcomm and Apple do not name the precision next to their NPU numbers, so a buyer cannot be sure two rows are measured the same way.

TOPS also ignore memory bandwidth and software. If a model is larger than the NPU's local memory, or needs an operation the NPU lacks, the work drops back to the CPU or GPU and the benefit is lost. Intel's page for its Series 3 parts lists the GPU at 122 TOPS INT8 and the NPU at 50, which shows how much of the total can sit outside the NPU.

Finally, TOPS describe peak compute, while many language-model tasks are limited by memory speed. A chip with a smaller TOPS figure and faster memory can produce text more quickly. Use TOPS to rule out chips below the bar a feature needs, not to rank those above it. Two chips with similar figures can behave very differently once the model, the software and the memory are the same for both.

What published benchmarks show#

Independent NPU data is thin. ComputerBase ran UL Procyon AI Computer Vision, integer format, on a set of chips shortly after the Snapdragon X Elite launched. The Core Ultra 7 165H has an older NPU, and the table shows its three engines side by side. The Snapdragon X Elite and the RTX 4070 laptop rows are there for scale, not as a head-to-head with the Intel chip, because the laptops, drivers and software stacks all differ.

Table
Chip and engineProcyon AI Computer Vision scoreBack end
Intel Core Ultra 7 165H, NPU358OpenVINO
Intel Core Ultra 7 165H, integrated GPU600OpenVINO
Intel Core Ultra 7 165H, CPU233OpenVINO
Qualcomm Snapdragon X Elite, NPU1742SNPE
Nvidia GeForce RTX 4070 Laptop, GPU2473TensorRT

Source: ComputerBase (opens in a new tab). Each vendor used its own software back end, so these are not like-for-like model runs.

ComputerBase also ran Geekbench ML, the predecessor of Geekbench AI. The test could not use the AMD NPU, so the table has no NPU column.

Table
ChipCPU scoreIntegrated GPU score (DirectML)
AMD Ryzen 9 8945HS31215332
Qualcomm Snapdragon X Elite30342316
Intel Core Ultra 7 165H28672749

Source: ComputerBase (opens in a new tab). The integrated GPUs ran through DirectML, while the CPU runs use each vendor's own path, so the columns are not directly comparable across vendors. The useful reading is within each row: on the AMD chip the integrated GPU is well ahead of the CPU, while on the Snapdragon the CPU is ahead of the GPU.

Vendor-run results exist as well. Intel reported 1.09 seconds to first token and 18.55 tokens/s for a Core Ultra 9 288V NPU on Llama 2 7B in MLPerf Client, and it described this as full NPU support in that benchmark. That is Intel's own run, so treat it as a maker claim and not as independent data.

Two lessons follow. The integrated GPU was ahead of the NPU on the older Intel chip, so a bigger NPU does not guarantee faster AI. And independent results for NPUs at 45 TOPS and above are still scarce. When independent numbers appear, check which framework and model they used.

The Copilot+ line and what it means#

Microsoft's developer guide sets 40 TOPS as the NPU level for many new Windows AI features. The 258V, the HX 370, the Snapdragon X chips and the 388H clear that line. The 265H, at 13 TOPS, does not. If a Copilot+ feature matters to you, check the NPU figure on the exact chip, not just the family name.

Most local AI tools that developers use, such as llama.cpp, Ollama or PyTorch, run on the CPU or GPU, and they do not need an NPU. NPU support is growing through frameworks such as ONNX Runtime and OpenVINO, but it is not yet the default path.

Desktop processors and the NPU#

Desktop chips mostly leave the NPU out. Intel lists a 13 TOPS NPU on the Core Ultra 9 285K (opens in a new tab), with 36 TOPS for the whole platform, well below the Copilot+ line. AMD's Ryzen 9 9950X page (opens in a new tab) lists no NPU at all. For a desktop AI build, the graphics card and the CPU's vector units matter far more than any NPU.

Phoronix found Ryzen 9000 chips ahead of Intel's Arrow Lake in CPU AI work and credits AVX-512, while Puget's MLPerf Client run on GPUs showed little CPU influence. The AI and ML CPU guide sorts the desktop chips on those results.

What this means when buying#

For a laptop, shortlist chips by the features you need. If you want Copilot+ features, look for 40 TOPS or more on the NPU. If you want to run models locally, look at the GPU and memory first, and treat the NPU as a bonus. Our AI laptop vs desktop for machine learning piece covers that trade-off, and the best Copilot+ PC guide lists machines that meet the line.

In India, laptops sell in many configurations under one model name, so check the chip, the memory size and the memory type on the maker's page for the exact model before you buy. Check whether the memory can be upgraded, because in many thin laptops it cannot.

For a desktop, pick the CPU by core count and AI test results, as in our AI workstation CPU guide, and spend on the GPU. The AI processors category lists every chip comparison on this site, including the 9950X vs 285K matchup.

Laptop availability in India lags launch dates abroad, so check which of these chips are actually sold through Amazon.in and the maker's Indian stores before you plan a purchase around one.

Frequently asked questions#

What does TOPS mean?

TOPS stands for trillions of operations per second. It is a peak figure for a stated precision, usually INT8, and real software rarely reaches it.

Is a higher NPU TOPS number always better?

No. Precision, memory and software support change real speed. ComputerBase measured the integrated GPU ahead of the NPU on an older Intel chip, so a higher TOPS figure does not guarantee a faster result in an app.

How many TOPS does Copilot+ need?

Microsoft's developer guide puts the NPU requirement for many new Windows AI features at 40 TOPS. Chips below that rating can still run the features that use the CPU or GPU, but some newer ones are limited to the NPU.

Do desktop CPUs have NPUs?

Some do. Intel lists 13 TOPS on the Core Ultra 9 285K. AMD's page shows no NPU on Ryzen 9000 desktop chips. Neither reaches the 40 TOPS Microsoft asks for, so desktop buyers should plan around the GPU.

Does an NPU speed up local LLMs?

Not necessarily. Most local LLM tools use the CPU or GPU, and the NPU helps mainly with low-power background AI tasks. For fast language-model speed, memory size and a good graphics card matter far more.

Which laptop chip has the highest NPU figure?

Qualcomm lists up to 85 TOPS for the Snapdragon X2 Elite family, and AMD lists 60 for the Ryzen AI 9 HX 475. The figures are not always stated at the same precision.

Sources & credits#

We have not tested these products ourselves. Specs come from the makers; benchmark figures come from the sources we name.

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.

What questions do readers ask next?

Back to top