NPU Device#

The Neural Processing Unit is a low-power hardware solution, introduced with the Intel® Core™ Ultra generation of CPUs (formerly known as Meteor Lake). It enables you to offload certain neural network computation tasks for more streamlined resource management.

NPU Plugin is now available through all relevant OpenVINO distribution channels.

Supported Platforms:
Host: Intel® Core™ Ultra series
NPU device: NPU 3720
OS: Ubuntu* 22.04 64-bit (with Linux kernel 6.6+), MS Windows* 11 64-bit (22H2, 23H2)

NPU Plugin needs an NPU Driver to be installed on the system to execute a model. Follow the instructions below to install the latest NPU drivers:

Note

Starting with the 2026.0 release, the compiler library is available in the OpenVINO package as a preview feature (Compiler-In-Plugin). The default compiler type remains Compiler-In-Driver (the compiler library included in the driver package). Users can override the default compiler selection by setting ov::intel_npu::compiler_type.

The plugin uses either the NPU compiler library included in the driver or the compiler library included in the OpenVINO package to convert the OpenVINO specific representation of the model into a proprietary format. The compiler performs platform specific optimizations in order to efficiently schedule the execution of layers and memory transactions on various NPU hardware submodules.

To use NPU for inference, pass the device name to the ov::Core::compile_model() method:

    core = ov.Core()
    compiled_model = core.compile_model(model, "NPU")

    ov::Core core;
    auto model = core.read_model("model.xml");
    auto compiled_model = core.compile_model(model, "NPU");

Model Caching#

Model Caching helps reduce application startup delays by exporting and reusing the compiled model automatically. The following two compilation-related metrics are crucial in this area:

First Ever Inference Latency (FEIL)
Measures all steps required to compile and execute a model on the device for the first time. It includes model compilation time, the time required to load and initialize the model on the device and the first inference execution.
First Inference Latency (FIL)
Measures the time required to load and initialize the pre-compiled model on the device and the first inference execution.

UMD Dynamic Model Caching#

UMD model caching is a solution enabled by default in the current NPU driver. It improves time to first inference (FIL) by storing the model in the cache after the compilation (included in FEIL), based on a hash key. The process may be summarized in three stages:

  1. UMD generates the key from the input IR model and build arguments

  2. UMD requests the DirectX Shader cache session to store the model with the computed key.

  3. All subsequent requests to compile the same IR model with the same arguments use the pre-compiled model, reading it from the cache instead of recompiling.

UMD Dynamic Model Caching can be bypassed for given model by setting boolean property ov::intel_npu::bypass_umd_caching (NPU_BYPASS_UMD_CACHE) to true at compilation. (default value is false)

OpenVINO Model Caching#

OpenVINO Model Caching is a common mechanism for all OpenVINO device plugins and can be enabled by setting the ov::cache_dir property. This way, the UMD model caching is automatically bypassed by the NPU plugin, which means the model will only be stored in the OpenVINO cache after compilation. When a cache hit occurs for subsequent compilation requests, the plugin will import the model instead of recompiling it.

For more details about OpenVINO model caching, see the Model Caching Overview.

Supported Inference Data Types
The NPU plugin supports the following data types as inference precision of internal primitives:
Floating-point data types: F32, F16
Quantized data types: U8 (quantized models may be INT8 or mixed FP16-INT8)
Computation precision for the HW is FP16.

For more details on how to get a quantized model, refer to the Model Optimization guide and NNCF tool quantization guide.

Supported Features and properties#

The NPU device is currently supported by AUTO inference modes (HETERO execution is partially supported, for certain models).

Supported Properties:

ov::device::id
ov::log::level
ov::hint::enable_cpu_pinning
ov::hint::inference_precision
ov::hint::model_priority
ov::hint::num_requests
ov::hint::performance_mode
ov::hint::execution_mode
ov::cache_dir
ov::compilation_num_threads
ov::enable_profiling
ov::workload_type
ov::intel_npu::compilation_mode_params
ov::intel_npu::compiler_dynamic_quantization
ov::intel_npu::qdq_optimization
ov::intel_npu::qdq_optimization_aggressive
ov::intel_npu::turbo
ov::intel_npu::platform
ov::intel_npu::tiles
ov::intel_npu::max_tiles
ov::intel_npu::bypass_umd_caching
ov::intel_npu::defer_weights_load
ov::intel_npu::run_inferences_sequentially
ov::intel_npu::disable_idle_memory_prunning
ov::intel_npu::compiler_type
ov::intel_npu::enable_strides_for

ov::supported_properties
ov::available_devices
ov::optimal_number_of_infer_requests
ov::range_for_async_infer_requests
ov::range_for_streams
ov::num_streams
ov::execution_devices
ov::device::architecture
ov::device::capabilities
ov::device::full_name
ov::device::uuid
ov::device::luid (windows only)
ov::device::pci_info
ov::device::gops
ov::device::type
ov::intel_npu::device_alloc_mem_size
ov::intel_npu::device_total_mem_size
ov::intel_npu::driver_version
ov::intel_npu::compiler_version

Note

The optimum number of inference requests returned by the plugin based on the performance mode is 4 for THROUGHPUT and 1 for LATENCY. The default mode for the NPU device is LATENCY.

ov::intel_npu::compilation_mode_params

ov::intel_npu::compilation_mode_params is an NPU-specific property that allows control of model compilation for NPU.

Note

The functionality is in experimental stage currently, can be a subject for deprecation and may be replaced with generic OV API in future OV releases.

Following configuration options are supported:

optimization-level

Defines an optimization effort hint to the compiler.

Value

Description

0

Reduced subset of optimization passes. May result in smaller compile time.

1

Default. Balanced performance/compile time.

2

Prioritize performance over compile time that may be an issue.

performance-hint-override

The LATENCY mode can be overridden by specifying ov::hint::performance_mode Has no effect for other ov::hint::PerformanceMode hints.

Value

Description

efficiency

Default. Balanced performance and power consumption.

latency

Prioritize performance over power efficiency.

Usage example:

map<str, str> config = {ov::intel_npu::compilation_mode_params.name(), ov::Any("optimization-level=1 performance-hint-override=latency")};

compile_model(model, config);

ov::intel_npu::turbo

The turbo mode, where available, provides a hint to the system to maintain the maximum NPU frequency and memory throughput within the platform TDP limits. The turbo mode is not recommended for sustainable workloads due to higher power consumption and potential impact on other compute resources.

core.set_property("NPU", ov::intel_npu::turbo(true));

or

core.compile_model(ov_model, "NPU", {ov::intel_npu::turbo(true)});

Note

NPU_TURBO usage may cause higher compile time, memory footprint, affect workload latency and compatibility issues with older NPU drivers

ov::intel_npu::max_tiles and ov::intel_npu::tiles

For on-device compilation, the plugin queries the driver for the available number of tiles and sets ov::intel_npu::max_tiles.

ov::intel_npu::max_tiles is a read-write property to allow users to set it during offline compilation.

Note that ov::intel_npu::max_tiles represents the maximum number of tiles available, but the compiler may target a lower number of tiles depending on other properties. Users can set ov::intel_npu::tiles to override the number of tiles selected by the compiler based on other properties.

Note

When setting ov::intel_npu::tiles, users must ensure that the value does not exceed ov::intel_npu::max_tiles. Any tile count other than 1 may impact cross-device compatibility if it is not explicitly validated against the target devices’s ov::intel_npu::max_tiles value.

ov::intel_npu::compiler_type

This property allows users to override the default compiler type selected by the plugin.

To use Compiler-In-Plugin whenever possible, users can set the property to PREFER_PLUGIN. This instructs the plugin to use the integrated compiler when all the following conditions are met:

  • The library is present

  • The compiler supports the current platform or there is no platform detected (offline compilation)

  • Compatibility is maintained between the current compiler version and all drivers released for the platform or there is no platform detected (offline compilation)

Note: On Meteor Lake (3720), when the property is set to PREFER_PLUGIN, the plugin will fall back to Compiler-in-Driver because the compiler library integrated in the plugin may not be compatible with driver versions lower than v2565. Users can set ov::intel_npu::compiler_type to PLUGIN to force Compiler-in-Plugin, but the blob will fail to execute on incompatible drivers.

If any condition is not met, the plugin automatically falls back to using Compiler-In-Driver. The compiler type used to compile a model can be queried from the resulting CompiledModel by reading the ov::intel_npu::compiler_type property.

Note

For on-device compilation with Compiler-In-Plugin, the plugin is responsible for querying the platform information from the driver and for passing the mandatory configs to the compiler (platform ID, available number of tiles, stepping information). Such compiled models are compatible with all the drivers released for that platform.

For offline compilation, users must explicitly set the ov::intel_npu::platform property to one of the supported values (see table above). Setting extra properties during offline compilation may result in compiled models that cannot be executed on SKUs with fewer resources or on drivers that do not support those features.

Example: Setting performance-hint-override=latency through ov::intel_npu::compilation_mode_params instructs the compiler to use all available resources for the given platform. If ov::intel_npu::max_tiles is not provided, the compiler falls back to a fixed lookup table embedded in the library to determine available resources, which might not be representative of all SKUs.

Limitations#

  • Currently, only models with static shapes are supported on NPU.

Import/Export:

Offline compilation and blob import is supported only for development purposes. Pre-compiled models (blobs) are not recommended to be used in production. Blob compatibility across different OpenVINO versions is not guaranteed.

Additional Resources#