NPU Explained: What It Is and What It Accelerates

An NPU (Neural Processing Unit) is a purpose-built processor for AI inference tasks, integrated directly into modern CPUs. Unlike a GPU, which handles general-purpose parallel computing, an NPU is optimised specifically for the matrix multiplication operations that underpin neural networks. The result is efficient, low-power inference for targeted AI tasks, without the heat, power draw, or cost of a discrete GPU.

What an NPU Is and How It Differs from CPU and GPU

The three processor types each have different strengths for AI workloads:

• CPU: General-purpose, sequential-first processing. Good at logic, branching, and low-latency single tasks. Slow at the massive parallel matrix operations AI models require. Available on every computer.
• GPU: Thousands of parallel cores designed originally for graphics. Excellent at the matrix multiplications used in AI training and inference. High power draw (100-400W for consumer cards). Expensive. Not present in most NAS.
• NPU: A dedicated accelerator for specific AI inference operations. Much lower power draw than a GPU (1-10W typically). Faster than CPU for targeted tasks. Cannot handle large general-purpose models. Integrated into modern CPUs.

An NPU is not a replacement for a GPU. It cannot run a 70B language model. What it can do is handle the AI features that vendors market as "AI NAS" features, such as photo face recognition, smart album generation, document OCR, and keyword search, far more efficiently than routing that work through the CPU.

How NPU Performance Is Measured: TOPS Explained

NPU performance is measured in TOPS: Tera Operations Per Second. One TOPS means the processor can perform one trillion AI operations per second. This metric specifically refers to 8-bit integer (INT8) operations, which is the standard precision for efficient inference.

To put the numbers in context:

• 1-5 TOPS: Basic image classification, simple object detection
• 10-20 TOPS: Real-time face recognition, photo categorisation, on-device voice processing
• 40+ TOPS: Threshold for Microsoft Copilot+ PC certification. Enables local Windows AI features.
• 100+ TOPS: Qualcomm Snapdragon X Elite range. Capable of running 7B language models at reasonable speeds on-NPU.

For NAS use, the NPU TOPS figure determines which vendor AI features run locally versus calling a cloud API. A 10 TOPS NPU handles Synology Photos face recognition locally. A 40 TOPS NPU handles Ollama's 3B-7B models meaningfully faster than CPU alone.

Which CPUs Have NPUs in 2026

NPU support in mainstream CPUs is recent. Here are the relevant families:

The key takeaway: the Intel Celeron processors that power most consumer NAS devices (QNAP TS-464, Synology DS925+, UGREEN DXP4800 Plus) have no NPU. AI features on these devices run on the CPU general cores, which is slower and less efficient than a dedicated NPU would be.

Which NAS Devices Have NPU Hardware in Australia

As of April 2026, NPU-equipped NAS devices available through Australian retail are limited to QNAP's workstation TVS-H series, which uses Intel Core Ultra processors.

Synology has not announced a consumer NAS with an integrated NPU as of early 2026. Synology Photos face recognition and AI search features run either on Synology's cloud infrastructure or via the host CPU. Synology's enterprise units use higher-end Intel Xeon processors, which also lack integrated NPUs.

UGREEN's AI marketing refers to software AI features in UGOS Pro (semantic photo search, document recognition) that leverage cloud APIs and CPU-side processing, not a dedicated NPU on the hardware.

What an NPU Actually Accelerates on a NAS

On QNAP's Core Ultra NAS devices, the Intel AI Boost NPU (~11 TOPS) accelerates these tasks meaningfully:

• AI photo recognition: Face detection, subject classification, scene recognition run faster and use less CPU headroom
• Document OCR (Paperless-ngx or QNAP Notes): Text extraction from scanned documents happens faster than CPU-only
• Small language model inference (1B-3B models): Models in this range see real throughput improvement on NPU versus CPU
• Video analytics: Object detection for surveillance applications, motion zone analysis
• Transcription (Whisper small/base): Audio-to-text conversion at faster-than-real-time speeds

At 11 TOPS, the Intel AI Boost NPU in current Meteor Lake chips is not fast enough to run a 7B model at interactive speeds. It will be noticeably faster than the CPU cores for these tasks, but a 7B Ollama model on an 11 TOPS NPU still generates tokens slowly. The Ryzen AI 300 (50+ TOPS) is a more capable tier for LLM inference.

What an NPU Cannot Do

Vendor marketing blurs the line between what an NPU handles and what a GPU handles. To be specific:

• Run large language models at interactive speeds: A 7B model on an 11 TOPS NPU will still be slow. A 13B+ model requires more compute than current embedded NPUs can handle at acceptable latency.
• Replace a GPU for image generation (Stable Diffusion): Stable Diffusion and ComfyUI require GPU-class VRAM and compute. No current NAS NPU can run these.
• General parallel compute (video encoding, 3D rendering): NPUs are specialised for AI matrix operations, not general-purpose parallel work.
• Run multiple AI workloads simultaneously: NPU resources are shared; running face recognition while also doing LLM inference will compete for the same limited TOPS.

NPU vs GPU: When Each Makes Sense

The choice between an NPU-equipped NAS and a GPU-capable NAS is a workload question:

• Use NPU if: your primary AI workloads are photo recognition, document OCR, transcription, surveillance object detection, or small model (1B-3B) inference. An NPU-equipped NAS handles these efficiently without adding a discrete GPU.
• Use GPU if: you need to run 7B+ language models at interactive speeds, generate images with Stable Diffusion, or serve inference to multiple concurrent users. A GPU is required; an NPU alone is not sufficient.
• Use CPU only if: your AI workloads are low-frequency (a few queries per day), quality matters more than speed, and hardware cost is a constraint. CPU inference on a 7B Q4 model is slow but workable for batch document processing.

Australian Context: NPU NAS Buying Considerations

If NPU acceleration is a requirement, the QNAP TVS-H series is the only current path in Australian retail. Availability is limited primarily to Scorptec, with some models on indent order. Confirm warehouse stock versus on-order before purchasing, as lead times for QNAP workstation units can be 2-4 weeks.

The price premium for NPU capability is significant. A QNAP TVS-H474 (entry TVS-H, Core Ultra i3) runs $2,400+ versus $600-900 for a QNAP TS-464 (Celeron, no NPU). Evaluate whether your AI workload justifies this premium. For most home users, the TS-464 running Ollama via CPU is adequate for the AI tasks they will actually use daily.

Mini-PCs with AMD Ryzen AI 300 (50 TOPS NPU) offer a higher-TOPS NPU at lower cost than the QNAP TVS-H series, but without the NAS storage architecture. The mini-PC vs NAS for local AI comparison covers this trade-off in detail.

Running a QNAP TVS-H with active AI inference adds measurable load to the system and electricity draw. Use the NAS power cost calculator with an estimate for continuous NPU workload when budgeting for ongoing running costs.

Related reading: our NAS buyer's guide.

Does my Synology NAS have an NPU?

No. As of April 2026, Synology does not sell consumer or prosumer NAS hardware with an integrated NPU. Synology's AI features (Synology Photos face recognition, smart search) run on the CPU or use Synology's cloud services. Synology has not announced an NPU-equipped NAS model for the consumer or prosumer market.

What is the difference between an NPU and a VPU?

A VPU (Vision Processing Unit) is a subset of NPU design optimised specifically for computer vision tasks: image classification, object detection, video analysis. Intel's older Movidius chips were VPUs. Modern Intel NPUs (AI Boost in Core Ultra) handle a broader range of AI workloads including language model inference, not just vision. In practice, the terms are often used interchangeably in marketing material. For NAS purposes, the TOPS rating matters more than the label.

Can Ollama use an NPU for acceleration?

Ollama supports CUDA (NVIDIA GPU) and ROCm (AMD GPU) acceleration natively. NPU support in Ollama depends on the backend; as of early 2026, Ollama's NPU support on Intel AI Boost is experimental and requires specific llama.cpp build flags. CPU inference remains the stable path on NPU-equipped hardware without a discrete GPU. This is expected to improve as the llama.cpp project matures its OpenVINO/IPEX-LLM backend support.

Is 11 TOPS enough for useful AI tasks on a NAS?

Yes, for targeted tasks. 11 TOPS (Intel AI Boost, Meteor Lake) is sufficient for AI photo recognition, document OCR, transcription with Whisper small/base, and object detection for surveillance. It is not sufficient for interactive 7B+ language model inference at useful speeds. Think of it as capable for ambient AI features, not a replacement for a GPU for conversational AI.

Will future Synology or UGREEN NAS have NPUs?

It is likely, though not confirmed. As Intel's Core Ultra platform becomes more widely adopted and AMD Ryzen AI series reaches embedded NAS form factors, NPU-equipped consumer NAS will become more common. UGREEN has shown roadmap interest in AI-capable hardware. Synology has been conservative with hardware changes historically. The next generation of prosumer NAS (2026-2027) is the likely timeframe for NPU-equipped options beyond QNAP.

How does an NPU compare to the AI features on Apple Silicon (M-series chips)?

Apple's M-series chips (M1, M2, M3, M4) have some of the most capable integrated NPUs available. M3 delivers ~18 TOPS from its NPU alone, with the full unified memory architecture enabling GPU inference on large models at speeds no NAS CPU can match. This is why Mac Mini M-series hardware has become popular for local AI setups. NAS NPUs in 2026 are meaningfully less capable than Apple Silicon for AI workloads, though the architecture serves the NAS use case (always-on, low power, storage-integrated) that Apple Silicon Mini cannot replace.