Best NPU for AI 2026: Intel vs Qualcomm vs AMD vs Apple

Short answer: in 2026 the best NPU is the one attached to the right memory, not the one with the biggest TOPS number. Qualcomm's Snapdragon X2 Elite is the fastest NPU on the spec sheet (80 TOPS, up to 85 on Elite Extreme), Apple's M5 Max is the best choice for large local LLMs because of unified memory capacity and bandwidth, and Intel Panther Lake or AMD Ryzen AI 400 are the picks if you need native x86. Every current part already clears Microsoft's 40-TOPS Copilot+ floor, so TOPS is no longer what separates them.

Which NPU is best for you?

Pick by the constraint you actually have, not by the TOPS headline:

• You want to run 30B-70B+ models locally → Apple M5 Max or M4 Max (128GB unified memory, the highest published bandwidth of any laptop part). Qualcomm X2 Elite Extreme now also reaches 128GB, which makes Windows-on-Arm viable for the first time.
• You need native x86 apps → Intel Panther Lake (Core Ultra 300) or AMD Ryzen AI 400. No emulation layer, no compatibility surprises.
• You want the highest NPU rating on the spec sheet → Qualcomm Snapdragon X2 Elite / Elite Extreme.
• You only want Windows Copilot+ features → literally any current NPU. The bar is 40 TOPS and every 2026 part clears it, so buy on screen, battery and price instead.
• You are not sure a laptop is the right machine at all → read the honest guide to running local AI on a laptop first; for many models a desktop GPU is still the cheaper answer.

---

What is an NPU, and why does it matter for local AI?

A Neural Processing Unit (NPU) is a processor block built for neural-network inference — principally the matrix multiplications that dominate AI workloads. Unlike a CPU (general purpose) or a GPU (wide parallel throughput), an NPU is tuned for:

• Matrix multiplication at the core of neural networks
• Low-power, always-on operation rather than peak throughput
• Inference without dedicated VRAM — it reads system/unified memory

That profile is what makes background AI features practical on battery: live captions, background blur and eye-contact correction, voice transcription, on-device search. It is also why an NPU is not a substitute for a discrete GPU on heavy generative work — see do I still need an NPU if I have a GPU? below.

The one hard threshold worth remembering is Microsoft's: a Copilot+ PC requires an NPU rated at 40 TOPS or higher, along with 16GB of RAM and 256GB of storage (Microsoft's Copilot+ PC requirements).

---

Is Intel Panther Lake (Core Ultra 300) the one to buy?

Intel's Panther Lake — Core Ultra Series 3, also branded Core Ultra 300 — is the successor to Lunar Lake and Intel's mainstream x86 AI-PC baseline for 2026. Announced at CES 2026, it is Intel's first AI PC platform on the Intel 18A process, with laptops shipping from late January 2026.

Specifications (Intel-published)

• NPU: NPU 5 architecture, up to 50 TOPS (INT8) on the top Core Ultra X9/X7 SKUs — clears the 40-TOPS Copilot+ bar
• Total platform AI: up to 180 TOPS (≈50 from the NPU plus ≈120 from the Xe3 Arc GPU)
• Process: Intel 18A, designed and manufactured in the US
• CPU: up to 16 cores (up to 4 Cougar Cove P-cores + 8 Darkmont E-cores + 4 low-power Darkmont E-cores)
• GPU: up to 12 Xe3 Arc cores with XMX units for AI acceleration
• Memory: LPDDR5X on-package, up to roughly 32GB

How it compares to Lunar Lake

The NPU itself barely moves: 48 TOPS (NPU 4) to about 50 TOPS (NPU 5) on the top SKUs. The bigger change is platform-wide — Intel's own CES 2026 claims are up to 60% better multithreaded CPU performance and up to 77% faster gaming versus Lunar Lake, with total platform AI nearly doubling to ~180 TOPS because of the Xe3 GPU. Those are Intel's figures, not independent test results. The OpenVINO and Windows ML developer paths are unchanged.

Caveat before you buy: 50 TOPS is the top-tier (Core Ultra X9/X7) number. Lower Core Ultra 5/3 Panther Lake SKUs ship with smaller GPU and AI configurations, so check the exact SKU on Intel's newsroom and ARK listings rather than assuming the flagship figure.

Best use cases

• 2026 x86 AI laptops that need the current Copilot+ baseline
• Native x86 compatibility plus the OpenVINO ecosystem
• Workloads that lean on the Xe3 GPU for the bulk of the ~180-TOPS platform total

---

How does Intel Lunar Lake (Core Ultra 200V) compare now?

Lunar Lake is still widely available and heavily discounted, which makes it the value option rather than the obsolete one.

Against Meteor Lake's NPU 3, rated at roughly 10 TOPS, Lunar Lake's 48-TOPS NPU 4 is about a 4.8x jump on the spec sheet (48 ÷ 10) — the single largest generational NPU increase Intel has shipped.

Architecture

• Unique feature: retains FP16 support, where AMD and Qualcomm NPUs are INT8-first
• Memory: LPDDR5X on-package, up to 32GB
• Integration: tightly coupled with the Xe2 GPU

Developer support

Primary SDK: OpenVINO

# OpenVINO NPU inference
from openvino import Core, compile_model

core = Core()
model = core.read_model("model.xml")
compiled = core.compile_model(model, "NPU")

# Run inference
result = compiled([input_tensor])

Framework support:

• torch.compile backend integration
• Keras 3.x backend support
• ONNX Runtime via the OpenVINO Execution Provider
• Windows ML automatic NPU selection

Intel was also first to land full NPU support in the MLPerf Client benchmark, which is the closest thing the category has to a neutral yardstick — see the benchmarks section below.

---

Is Qualcomm Snapdragon X2 Elite the fastest NPU?

On vendor-published TOPS, yes — by a wide margin.

The X2 series nearly doubles the rated NPU throughput from 45 to 80 TOPS, with top Elite Extreme SKUs rated up to 85. Notebooks ship through the first half of 2026.

Architecture

• Process: TSMC 3nm (X2 series)
• Total platform AI: vendor-quoted at 100+ TOPS across CPU, GPU, NPU and micro NPU
• Micro NPU: always-on sensing for human presence detection
• Memory: up to 128GB on-package LPDDR5X-9523 with a 12-channel bus (228 GB/s) on X2 Elite Extreme — a major jump from the early 48GB leak figures, and enough to hold 70B-class local LLMs on a Windows-on-Arm machine for the first time

That memory change is the genuinely important one. Until 2026 the answer to "can a Windows laptop run a 70B model?" was simply no. Now it is yes, at lower bandwidth than Apple. If you are sizing a machine around a specific model, our VRAM and unified-memory requirements guide maps parameter count and quantisation to the memory you actually need.

Developer support

Primary SDK: AI Engine Direct

// Qualcomm AI Engine Direct
#include "QnnInterface.h"

// Load model
Qnn_ModelHandle_t model;
QnnModel_create(modelPath, &model);

// Execute inference on NPU
QnnModel_executeGraphs(model, inputs, outputs);

Framework support:

• ONNX Runtime via the QNN Execution Provider
• Windows ML native integration
• LiteRT (Google) support in progress
• QAI AppBuilder for simplified deployment

The trade-off

Snapdragon X2 is Arm, so x64 Windows applications run under emulation. That has improved a great deal but it is still the reason a lot of buyers choose Intel or AMD instead. Arm-based Snapdragon X laptops are also consistently reviewed as the battery-life leaders in the Windows category — but published battery figures swing enormously with screen brightness and test workload, so check a review of the exact model you are considering rather than trusting a single headline number.

---

What is AMD XDNA, and how does Ryzen AI 400 perform?

The flagship Ryzen AI 9 HX 475 pairs 12 Zen 5 cores with the 60-TOPS XDNA 2 NPU and RDNA 3.5 graphics. Gorgon Point is an iterative refresh of Strix Point — same Zen 5, RDNA 3.5 and XDNA 2 IP at higher clocks — not a new NPU generation, so the "XDNA 2+" label in some early coverage is informal. AMD's own materials call it XDNA 2.

Architecture

AMD XDNA descends from Xilinx technology:

• Design: spatially arranged AI Engine tiles
• Cores: VLIW + SIMD vector cores for matrix operations
• Memory: LPDDR5X-8533 support
• Integration: Zen 5 CPU + RDNA 3.5 GPU + XDNA 2 NPU

Developer support

Primary SDK: Ryzen AI Software

# AMD Vitis AI with ONNX Runtime
import onnxruntime as ort

# Create session with Vitis AI EP (auto NPU/CPU partitioning)
sess = ort.InferenceSession(
    "model.onnx",
    providers=["VitisAIExecutionProvider", "CPUExecutionProvider"]
)

# Run inference
result = sess.run(None, {"input": data})

Framework support:

• Vitis AI Execution Provider for ONNX Runtime
• AMD Quark quantizer (PyTorch and ONNX)
• Windows ML integration
• Supported precisions: INT8, BF16, FP32 (auto-converted to BF16)

ROCm status

AMD's stated position at CES 2026 was that it is focused on enabling the Windows path — Windows ML access plus continued polish of the Vitis libraries — rather than exposing the NPU directly through ROCm. Recent ROCm releases add support for Ryzen AI Max ("Halo") systems, but direct NPU programming through ROCm is still not available. In practice that means ONNX Runtime and Windows ML are the routes that work today.

Best use cases

• Full x86-64 compatibility with no emulation
• Windows gaming plus AI workflows on one machine
• Enterprise deployments that require x86
• Buyers who expect ROCm NPU support to arrive eventually

---

How does Apple's Neural Engine (M4 and M5) compare?

Apple is the outlier: the lowest published TOPS in this comparison and, for local LLMs, still the best machine. That is not a contradiction — it is what happens when memory bandwidth, not compute, is the binding constraint.

M4 family (Apple-published specs)

The M4 Neural Engine is rated at roughly 2.1x the M3's 18 TOPS (38 ÷ 18). Apple's own stated comparison is that M4's Neural Engine is 60x faster than the A11 Bionic of 2017 — an Apple marketing figure, quoted here as such.

M5 family: the architecture actually changed

The M-series moved on while most NPU comparisons were still quoting M4. The Apple M5 launched October 15, 2025, followed by M5 Pro and M5 Max on March 3, 2026 in the new MacBook Pro. The headline change is architectural: every GPU core now contains a dedicated Neural Accelerator, and Apple routes most on-device AI through the GPU rather than through the Neural Engine. Apple's claim is over 4x the peak GPU compute for AI versus M4, with prompt processing substantially faster on M5 Max — again, Apple's numbers, not independent measurements.

Honest note: Apple no longer publishes a Neural Engine TOPS figure for M5, so any "M5 = X TOPS" claim you see elsewhere is somebody's estimate, not an Apple specification. Current configurations and bandwidth figures are on Apple's MacBook Pro tech specs. For local-LLM buyers the numbers that matter are unchanged: unified memory capacity and bandwidth.

Developer support

Primary SDK: Core ML

// Core ML inference
import CoreML

let model = try! MyModel(configuration: MLModelConfiguration())
let input = MyModelInput(data: inputData)
let output = try! model.prediction(input: input)

MLX framework (open source):

# MLX for Apple Silicon
import mlx.core as mx
import mlx.nn as nn

# Arrays live in unified memory - no transfer needed
x = mx.array([1, 2, 3])
model = nn.Linear(input_dims, output_dims)
output = model(x)

Why Apple still wins for LLMs

For LLM decoding, memory bandwidth governs tokens per second once the weights fit in memory — every generated token has to stream the active weights out of RAM. As of 2026 the X2 Elite Extreme finally matches Apple on capacity (both 128GB), but Apple keeps a roughly 2.7x bandwidth lead (614 ÷ 228). Intel remains the most memory-constrained option in the group.

If you are weighing a Mac specifically, the Apple M4 for AI guide goes deeper on inference behaviour, and the Apple Silicon AI buying guide covers which configuration is worth the money.

---

What actually changed in 2026?

If you are reading an NPU comparison written in early 2026, three things are now out of date — and all three change buying advice:

1. Qualcomm closed the memory gap. The Snapdragon X2 Elite Extreme ships with up to 128GB of on-package LPDDR5X. Early coverage said 48GB. A Windows-on-Arm laptop can finally hold a 70B-class model in memory, something only Apple could do before.
2. Apple stopped quoting Neural Engine TOPS. With the M5 family, Apple moved AI compute into per-core GPU Neural Accelerators and now reports relative speedups instead of a TOPS number. Treat any "M5 = N TOPS" figure as an estimate.
3. Intel shipped Panther Lake. Core Ultra 300 (NPU 5, up to 50 TOPS, Intel 18A) replaced Lunar Lake as Intel's mainstream AI-PC baseline in January 2026 — but Intel still trails on standalone NPU TOPS and, more importantly, on on-package memory.

The practical takeaway has not moved: for running local models, RAM and bandwidth decide the outcome, not the NPU TOPS headline. Once your hardware is sorted, Ollama, LM Studio and Jan are the easiest ways to put it to work, and running an LLM on your laptop's NPU covers what actually executes on the NPU today versus what quietly falls back to CPU.

---

How do these NPUs compare on real benchmarks?

This is the part where most NPU comparison articles — including earlier versions of this one — start quoting seconds-per-image and tokens-per-second figures with no source attached. We have removed those numbers rather than repeat them, because we do not own this hardware and could not trace them to a primary result.

What we can tell you is where to get numbers you can actually verify, and how to read them:

Three things to watch for when you read any NPU benchmark:

• Which execution provider ran it. The same model on the same laptop can post very different scores through OpenVINO, QNN, Vitis AI or a generic CPU fallback. A "slow NPU" result is often a model that never reached the NPU at all.
• Which precision. NPU figures are usually INT8. Comparing an INT8 NPU score to an FP16 GPU score is not a comparison.
• NPU versus iGPU on the same chip. For image generation in particular, the integrated GPU is frequently the faster path while the NPU is the cooler and quieter one. Reviews that only publish the NPU number can make a chip look worse than it is in practice.

For heavy generative work — Stable Diffusion at speed, or anything above a 13B LLM — a discrete GPU still beats every NPU in this comparison. The NPU's job is efficient, always-on inference, not peak throughput. Our GPU VRAM requirements guide is the right reference if that is the workload you actually have.

---

Which NPU has the best developer tools?

Intel and Apple have the most mature documentation — comprehensive OpenVINO guides and Hugging Face integration on one side, Core ML docs, WWDC sessions and MLX tutorials on the other. Qualcomm's AI Engine Direct documentation and AMD's Ryzen AI documentation are both improving quickly but still assume more prior knowledge. All four support the ONNX model format, so portability is genuinely better than it was two years ago.

Model conversion workflow

PyTorch Model → ONNX Export → Quantization → Platform Deploy

Intel:     PyTorch → ONNX → OpenVINO MO → .xml/.bin
Qualcomm:  PyTorch → ONNX → QNN Converter → .qnn
AMD:       PyTorch → ONNX → AMD Quark → .onnx (quantized)
Apple:     PyTorch → coremltools → .mlpackage

If your day job is coding rather than model deployment, note that most AI coding assistants still run in the cloud or on a discrete GPU — the NPU is not yet the thing accelerating your editor.

---

Which NPU should you buy?

Qualcomm Snapdragon X2 Elite

Choose if you need: the highest rated NPU throughput (80-85 TOPS), long battery life on Windows, an Arm-native Windows 11 machine, or 128GB of memory in a Windows laptop.

Trade-offs: x64 applications run under emulation, and the software ecosystem is still smaller than x86.

Intel Panther Lake (Core Ultra 300)

Choose if you need: the current x86 Copilot+ baseline, native application compatibility, the OpenVINO ecosystem, or a machine where the Xe3 GPU does the heavy AI lifting.

Trade-offs: lower standalone NPU TOPS than Qualcomm or AMD, and roughly 32GB of on-package memory is a hard ceiling for local LLMs.

AMD Ryzen AI 400

Choose if you need: full x86-64 native compatibility, gaming and AI on the same machine, or an enterprise x86 deployment.

Trade-offs: the NPU is not the fast path for image generation, and ROCm NPU support is still not shipping.

Apple M5 Max (or M4 Max)

Choose if you need: maximum unified memory (128GB), the highest published memory bandwidth (up to 614 GB/s on M5 Max, 546 GB/s on M4 Max), large local LLMs, or macOS creative workflows.

Trade-offs: macOS only, no Windows Copilot+ features, and no published Neural Engine TOPS number to compare against.

---

Which NPU is best for each use case?

---

What is coming in 2026-2027?

Trends to watch

1. Memory capacity is the real race — it matters more than raw TOPS for anyone running LLMs
2. NPU programming is maturing — ROCm and MLX are both closing gaps
3. TOPS is becoming a commodity — when every part clears the Copilot+ bar, the number stops differentiating
4. Software decides outcomes — execution-provider support now moves real-world speed more than silicon does

---

Common questions about NPUs

Which NPU has the highest TOPS in 2026?

On vendor-published figures, Qualcomm Snapdragon X2 Elite leads at 80 TOPS from its 6th-generation Hexagon NPU (NPU6), with X2 Elite Extreme SKUs rated up to 85. AMD Ryzen AI 400 "Gorgon Point" follows at 60 TOPS (XDNA 2), Intel Panther Lake reaches up to 50 TOPS (NPU 5), and Lunar Lake sits at 48. Apple no longer publishes a Neural Engine TOPS figure for M5. TOPS alone does not decide real-world performance — memory bandwidth, precision support and framework coverage matter at least as much.

What are Windows Copilot+ PCs and what NPU do they need?

Copilot+ PCs are Microsoft's AI-enabled Windows machines. The published requirement is an NPU rated at 40 TOPS or higher, plus 16GB of RAM and 256GB of storage. Features include Recall, Click to Do, Live Captions with translation and Windows Studio Effects. Intel Core Ultra 200V and 300, Qualcomm Snapdragon X and X2, and AMD Ryzen AI 300 and 400 all qualify. Macs do not run Windows Copilot+ features at all.

Which NPU is best for running local LLMs?

The one attached to the most memory at the highest bandwidth. Apple M4 Max, M5 Max and Qualcomm Snapdragon X2 Elite Extreme all reach 128GB of unified memory, which is enough for 70B-class models at typical quantisations. Apple keeps a bandwidth lead (up to 614 GB/s on M5 Max versus 228 GB/s on X2 Elite Extreme), and bandwidth is what governs generation speed once a model fits. Intel Panther Lake is the most constrained at roughly 32GB on-package. NPU TOPS is close to irrelevant here — model size is limited by RAM, and speed by bandwidth.

How does Intel's NPU compare to Qualcomm Hexagon?

Intel's NPU 4 (Lunar Lake) is rated 48 TOPS and NPU 5 (Panther Lake) up to 50, versus 80-85 TOPS for Qualcomm's Hexagon NPU6. Intel retains FP16 support, which is unusual among x86 NPUs, and has the strongest OpenVINO tooling. Qualcomm has the higher rating and the Arm power profile. The deciding factor for most buyers is not the NPU at all — it is whether you need native x86 applications (Intel) or are happy with emulation for the sake of battery life (Qualcomm).

What is AMD XDNA and how does it perform?

AMD XDNA is AMD's NPU architecture, inherited from Xilinx, built around spatially arranged AI Engine tiles with VLIW plus SIMD vector cores. XDNA 2 in Ryzen AI 300 "Strix Point" is rated 50 TOPS; Ryzen AI 400 "Gorgon Point" uses the same XDNA 2 NPU at higher clocks for 60 TOPS — an iterative refresh, not a new generation. AMD NPUs integrate with Vitis AI and ONNX Runtime. The key advantage is full x86-64 compatibility with no emulation; the key gap is that direct NPU programming through ROCm is still unavailable.

How does Apple's Neural Engine compare to Windows NPUs?

Apple's 16-core Neural Engine in M4 is rated around 38 TOPS, lower than Windows competitors on paper, but it comes with 128GB of unified memory and 546 GB/s of bandwidth on M4 Max plus tight Core ML and MLX integration. With M5, Apple moved most AI compute into per-core GPU Neural Accelerators and stopped quoting a Neural Engine TOPS number; M5 Max raises bandwidth to up to 614 GB/s. For LLMs, that bandwidth advantage matters more than raw TOPS. The limitation is macOS only, with no Windows Copilot+ features.

Which NPU is most power efficient for battery life?

Arm-based Snapdragon X and X2 laptops are consistently reviewed as the battery-life leaders in the Windows category, and Apple Silicon leads on macOS. Beyond that ordering, be sceptical of specific hour counts: published battery figures vary enormously with screen brightness, panel type, workload and review methodology, so a number measured on one laptop tells you very little about another. Check a review of the exact configuration you intend to buy. What is reliably true is that running a workload on the NPU rather than the CPU or GPU draws less power for the same task — that is the entire reason the block exists.

What developer tools and SDKs are available for each NPU?

Intel: OpenVINO (mature, with torch.compile and Keras backends) plus ONNX Runtime. Qualcomm: AI Engine Direct SDK, QAI AppBuilder and ONNX Runtime via the QNN execution provider. AMD: Ryzen AI Software, Vitis AI, the AMD Quark quantizer and ONNX Runtime via the Vitis AI execution provider. Apple: Core ML with excellent Xcode integration, plus the open-source MLX array framework. All four support ONNX. Intel and Apple have the most complete documentation.

Which NPU is best for Stable Diffusion image generation?

Honest answer: not one we can rank for you with numbers, because we have not run these parts. What is worth knowing before you buy is that on most Windows laptops the integrated GPU, not the NPU, is the faster path for diffusion models — the NPU is the quieter, cooler, lower-power path. Reviews that publish only the NPU figure understate what a chip can do. If image generation at speed is your main workload, a discrete GPU will beat every laptop NPU in this comparison by a wide margin, and you should be shopping on VRAM instead.

Do I need an NPU if I already have a powerful GPU?

They serve different purposes. GPUs win on training and high-throughput inference with dedicated VRAM. NPUs win on always-on, low-power inference that runs in the background without flattening your battery or spinning up fans, and they are required for Windows Copilot+ features. If you already own a capable discrete GPU, the NPU is not going to speed up your heavy generative work — treat it as a battery-life and background-features feature, not a performance upgrade.

What is the minimum TOPS needed for useful local AI?

The only hard, published threshold is Microsoft's: 40 TOPS for Windows Copilot+ features. Everything below that is workload-dependent rather than a fixed number — small vision models such as object detection and image classification have run acceptably on the 10-16 TOPS NPUs of 2023-era chips, while anything generative is bounded by memory long before it is bounded by TOPS. If your goal is running LLMs locally, size the machine on RAM and bandwidth first and treat the NPU rating as a tie-breaker.

---

Key takeaways

1. Qualcomm X2 leads on the spec sheet (80-85 TOPS) and on the Windows battery-life reputation
2. Apple wins for LLMs on bandwidth — M5 Max (128GB / up to 614 GB/s) and M4 Max (128GB / 546 GB/s); X2 Elite Extreme now matches the capacity but at 228 GB/s
3. AMD XDNA offers the best x86 compatibility for combined gaming and AI machines
4. Intel's OpenVINO is the most mature developer ecosystem
5. 40 TOPS is the Copilot+ minimum — every current part exceeds it, so TOPS no longer differentiates
6. Memory bandwidth matters more than TOPS for LLM inference
7. All four vendors support ONNX, so model portability keeps improving

---

Next steps

1. Check memory and VRAM requirements before choosing between NPU, iGPU and discrete GPU
2. See what actually runs on a laptop NPU versus what falls back to the CPU
3. Read the honest laptop local-AI guide if you are deciding between a laptop and a desktop
4. Set up local LLMs with Ollama, LM Studio or Jan
5. Compare Apple Silicon for AI in depth
6. Understand MoE models, which benefit disproportionately from limited-bandwidth hardware

---

NPUs have gone from niche accelerators to a standard part of every laptop. Whether you prioritise battery life (Qualcomm), memory capacity and bandwidth (Apple), x86 compatibility (AMD), or developer ecosystem (Intel), there is a part suited to your workflow — but in 2026 the NPU rating is rarely the number that should decide the purchase.