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 from other devices, for more streamlined resource management.

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

Supported Platforms:
Host: Intel® Core™ Ultra (former Meteor Lake)
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 for both compiling and executing a model. Follow the instructions below to install the latest NPU drivers:

The plugin uses the graph extension API exposed by the driver to convert the OpenVINO specific representation of the model into a proprietary format. The compiler included in the user mode driver (UMD) performs platform specific optimizations in order to efficiently schedule the execution of network 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).

The NPU support in OpenVINO is still under active development and may offer a limited set of supported OpenVINO features.

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::turbo
ov::intel_npu::tiles
ov::intel_npu::max_tiles
ov::intel_npu::bypass_umd_caching
ov::intel_npu::defer_weights_load

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::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

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);

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)});

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

the max_tiles property is read-write to enable compiling models off-device. When on NPU, max_tiles will return the number of tiles the device has. Setting the number of tiles to compile for (via intel_npu::tiles), when on device, must be preceded by reading intel_npu::max_tiles first, to make sure that ov::intel_npu::tiles <= ov::intel_npu::max_tiles to avoid exceptions from the compiler.

Note

ov::intel_npu::tiles overrides the default number of tiles selected by the compiler based on performance hints (ov::hint::performance_mode). Any tile number other than 1 may be a problem for cross platform compatibility, if not tested explicitly versus the max_tiles value.

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 / NPU Driver versions is not guaranteed.

Additional Resources#