Performance

Authors: Mingyu Kim, Vladimir Paramuzov, Nico Galoppo

Intel’s newest GPUs, such as Intel® Data Center GPU Flex Series, and Intel® Arc™ GPU, introduce a range of new hardware features that benefit AI workloads. Starting with the 2022.3 release, OpenVINO™ can take advantage of two newly introduced hardware features: XMX (Xe Matrix Extension) and parallel stream execution. This article explains what those features are and how you can check whether they are enabled in your environment. We also show how to benefit from them with OpenVINO, and the performance impact of doing so.

What is XMX (Xe Matrix Extension)?

XMX is a hardware acceleration for matrix multiplication on the newest Intel™ GPUs. Given the same number of Xe Cores, XMX technology provides 4-8x more multiplication capacity at the same precision [1]. OpenVINO, powered by OneDNN, can take advantage of XMX hardware by accelerating int8 and fp16 inference. It brings performance gains in compute-intensive deep learning primitives such as convolution and matrix multiplication.

Under the hood, XMX is a well-known hardware architecture called a systolic array. Systolic arrays increase computational capacity without increasing memory (or register) access. The magic happens by pipelining multiple computations with a single data access, as opposed to the traditional fetch-compute-store pipeline. It is implemented by connecting multiple computation nodes in series. Data is fed into the front, goes through several steps of multiplication-add, and finally is stored back to memory.

How to check whether you have XMX?

You can check whether your GPU hardware (and software stack) supports XMX with OpenVINO™’s hello_query_device sample. When you run the sample application, it lists all detected inference devices along with its properties. You can check for XMX support by looking at the OPTIMIZATION_CAPABILITIES property and checking for the GPU_HW_MATMUL value.

In the listing below you can see that our system has two GPU devices for inference, and only GPU.1 has XMX support.

$ ./hello_query_device
[ INFO ] GPU.0
[ INFO ]        SUPPORTED_PROPERTIES: 
[ INFO ]                Immutable: OPTIMIZATION_CAPABILITIES : FP32 BIN FP16 INT8      
# XMX is not supported
[ INFO ] GPU.1
[ INFO ]        SUPPORTED_PROPERTIES: 
[ INFO ]                Immutable: OPTIMIZATION_CAPABILITIES : FP32 BIN FP16 INT8 GPU_HW_MATMUL    
# XMX is supported

‍

As mentioned, XMX provides a way to get significantly more compute capacity on a GPU. The next feature doesn’t provide more capacity, but it allows ways to use that capacity more efficiently.

What is parallel execution of multiple streams?

Another improvement of Intel®’s discrete GPUs is to process multiple compute streams in parallel. Certain deep learning inference workloads are too small to fill all hardware compute resources of a given GPU. In such a case it is beneficial to run multiple compute streams (or inference requests) in parallel, such that the GPU hardware has more work to process at any given point in time. With parallel execution of multiple streams, Intel GPUs can increase hardware efficiency.

How to check for parallel execution support?

As of the OpenVINO 2022.3 release, there is only an indirect way to query how many streams your GPU can process in parallel. In the next release it will be possible to query the range of streams using the ov::range_for_streams property query and the hello_query_device_sample. Meanwhile, one can use the benchmark_app to report the default number of streams (NUM_STREAMS). If the GPU does not support parallel stream execution, NUM_STREAMS will be 2. If the GPU does support it, NUM_STREAMS will be larger than 2. The benchmark_app log below shows that GPU.1 supports 4-stream parallel execution.

$ ./benchmark_app -d GPU.0 -m resnet-50.xml -t 1 --hint none
[ INFO ]   NUM_STREAMS: 2      # Single-stream execution is supported
…
$ ./benchmark_app -d GPU.1 -m resnet-50.xml -t 1 --hint none
[ INFO ]   NUM_STREAMS: 4      # 4-stream execution is supported
…

‍

However, it depends on application usage

Parallel stream execution can bring significant performance benefit, but only when used appropriately by the application. It will bring good performance gain if the application can run multiple independent inference requests in parallel, whether from single process or multiple processes. On the other hand, if there is no opportunity for parallel execution of multiple inference requests, then there is no gain to be had from multi-stream hardware execution.

Demonstration of performance tuning through benchmark_app

DISCLAIMER: The performance may vary depending on the system and usage.

OpenVINO benchmark_app is a very handy tool to analyze performance in various conditions. Here we’ll show the performance trend for an Intel® discrete GPU with XMX and four parallel hardware execution streams.

The performance was measured on a pre-production version of the Intel® Arc™ A770 Limited Edition GPU with 16 GiB of memory. The host system is a 12th Gen Intel(R) Core(TM) i9-12900K with 64GiB of RAM (4 DDR4-2667 modules) running Ubuntu OS 20.04.5 LTS with Linux kernel 5.15.47.

Performance comparison with high-level performance hints

Even though all supported devices in OpenVINO™ offer low-level performance settings, utilizing them is not recommended outside of very few cases. The preferred way to configure performance in OpenVINO Runtime is using performance hints. This is a future-proof solution fully compatible with the automatic device selection inference mode and designed with portability in mind.

OpenVINO benchmark_app exposes the high-level performance hints with the performance hint option for easy configuration of best latency and throughput. In short, latency mode picks the optimal configuration for low latency with the cost of low throughput, and throughput mode picks the optimal configuration for high throughput with the cost of high latency.

The table below shows throughput for various combinations of execution configuration for resnet-50.

‍

Throughput mode is achieving much higher FPS compared to latency mode because inference happens with higher batch size and parallel stream execution.  You can also see that, in throughput mode, the throughput with fp16 is 5.4x higher than with fp32 due to the use of XMX.

In the experiments below we manually explore different configurations of the performance parameters for demonstration purposes; It is generally not recommended to tune manually. Once the optimal parameters are known, they can be applied in production.

Performance gain from XMX

Performance gain from XMX can be observed by comparing int8/fp16 against fp32 performance because OpenVINO does not provide an option to turn XMX off. Since fp32 computations are not executed by the XMX hardware pipe, but rather by the less efficient fetch-compute-store pipe, you can see that the performance gap between fp32 and fp16 is much larger than the expected factor of two.

We choose a batch size of 64 to demonstrate the best case performance gain. When the batch size is small, the performance difference is not always as prominent since the workload could become too small for the GPU.

$ ./benchmark_app -d GPU.1 -m resnet-50-fp.xml -t 10 --hint none --nstreams 4 -b 64 --infer_precision f32 | grep Throughput
[ INFO ] Throughput:          1076.22 FPS 
$ ./benchmark_app -d GPU.1 -m resnet-50-fp.xml -t 10 --hint none --nstreams 4 -b 64 --infer_precision f16 | grep Throughput
[ INFO ] Throughput:          5915.62 FPS
$ ./benchmark_app -d GPU.1 -m resnet-50-int8.xml -t 10 --hint none --nstreams 4 -b 64 | grep Throughput
[ INFO ] Throughput:          12270.12 FPS

‍

As you can see from the execution log, fp16 runs ~5.49x faster than fp32. Int8 throughput is ~2.07x higher than fp16. The difference between fp16 and fp32 is due to fp16 acceleration from XMX while fp32 is not using XMX. The performance gain of int8 over fp16 is 2.07x because both are accelerated with XMX.

Performance gain from parallel stream execution

You can see from the log below that performance goes up as we have more streams up to 4. It is because the GPU can handle 4 streams in parallel.

$./benchmark_app -d GPU.1 -m resnet-50-int8.xml -t 10 --hint none --nstreams 1 -b 64 | grep Throughput
[ INFO ] Throughput:          8593.92 FPS
$./benchmark_app -d GPU.1 -m resnet-50-int8.xml -t 10 --hint none --nstreams 2 -b 64 | grep Throughput
[ INFO ] Throughput:          10610.98 FPS
$./benchmark_app -d GPU.1 -m resnet-50-int8.xml -t 10 --hint none --nstreams 4 -b 64 | grep Throughput
[ INFO ] Throughput:          12246.29 FPS
$./benchmark_app -d GPU.1 -m resnet-50-int8.xml -t 10 --hint none --nstreams 8 -b 64 | grep Throughput
[ INFO ] Throughput:          12150.30 FPS

‍

Note that if the inference workload is large enough, more streams might not bring much or any performance gain. For example, when increasing the batch size, throughput may saturate earlier than at 4 streams.

How to take advantage the improvements in your application

For XMX, all you need to do is run your int8 or fp16 model with the OpenVINO™ Runtime version 2022.3 or above. If the model is fp32(single precision), it will not be accelerated by XMX. To quantize a model and create an OpenVINO int8 IR, please refer to Quantizing Models Post-training. To create an OpenVINO fp16 IR from a fp32 floating-point model, please refer to Compressing a Model to FP16 page.

For parallel stream execution, you can set throughput hint as described in Optimizing for Throughput. It will automatically set the number of parallel streams with best number.

Conclusion

In this article, we introduced two key features of Intel®’s discrete GPUs: XMX and parallel stream execution. Most int8/fp16 deep learning networks can benefit from the XMX engine with no additional configuration. When properly configured by the application, parallel stream execution can bring significant performance gains too!

‍

[1] In the Xe-HPG architecture, the XMX delivers 256 INT8 ops per clock (DPAS), while the (non-systolic) Xe Core vector engine delivers 64 INT8 ops per clock – a 4x throughput increase [reference]. In the Xe-HPC architecture, the XMX systolic array depth has been increased to 8 and delivers 4096 FP16 ops per clock, while the (non-systolic) Xe Core vector engine delivers 512 FP16 ops per clock – a 8x throughput increase [reference].

‍

Notices & Disclaimers

​Performance varies by use, configuration and other factors. Learn more at www.Intel.com/PerformanceIndex​​.

Performance results are based on testing as of dates shown in configurations and may not reflect all publicly available ​updates.  See backup for configuration details.  No product or component can be absolutely secure.​​

​​​​See backup for configuration details.  For more complete information about performance and benchmark results, visit www.intel.com/benchmarks

© Intel Corporation.  Intel, the Intel logo, and other Intel marks are trademarks of Intel Corporation or its subsidiaries.  Other names and brands may be claimed as the property of others​.​​

Starting with the 2022.3 release, OpenVINO Model Server (OVMS) provides a C-API that allows OVMS to be linked directly into a C/C++ application as a dynamic library. Existing AI applications can leverage serving functionalities while running inference locally without networking latency overhead.  

The ability to bypass gRPC/REST endpoints and send input data directly from in-process memory creates new opportunities to use OpenVINO locally while maintaining the benefits of model serving. For example, we can combine the benefits of using OpenVINO Runtime with model configuration, version management and support for both local and cloud model storage.

‍

OpenVINO Model Server is typically started as a separate process or run in a container where the client application communicates over a network connection. Now, as you can see above, it is possible to link the model server as a shared library inside the client application and use the internal C API to execute internal inference methods.

We demonstrate the concept in a simple example below and show the impact on latency.

‍

Example C-API Usage

NOTE: complete end to end inference demonstration via C-API with example app can be found here: https://docs.openvino.ai/latest/ovms_demo_capi_inference_demo.html  

To start using the Model Server C-API, we need to prepare a model and configuration file. Download an example dummy model from our GitHub repo and prepare a config.json file to serve this model. “Dummy” model adds value 1 to all numbers inside an input.

‍

Download Model

wget https://github.com/openvinotoolkit/model_server/raw/main/src/test/dummy/1/dummy.{xml,bin} -P models/dummy/1

‍

Create Config File

{ 
    "model_config_list": [ 
        {"config": { 
                "name": "dummy", 
                "base_path": "./models/dummy"}} 
    ] 
} 

‍

Get libovms_shared.so

Next, download and unpack the OVMS library. The library can be obtained from GitHub release page. There are 2 packages – one for Ubuntu 20 and one for RedHat 8.7. There is also documentation showing how to build the library from source. For purpose of this demo, we will use the Ubuntu version:

wget https://github.com/openvinotoolkit/model_server/releases/download/v2022.3/ovms_ubuntu.tar.gz && tar -xvf ovms_ubuntu.tar.gz

‍

Start Server

To start the server, use ServerStartFromConfigurationFile. There are many options, all of which are documented in the header file. Let’s launch the server with configuration file and optional log level error:

    OVMS_ServerSettings* serverSettings; 
    OVMS_ModelsSettings* modelsSettings; 
    OVMS_Server* srv; 
    OVMS_ServerSettingsNew(&serverSettings); 
    OVMS_ModelsSettingsNew(&modelsSettings); 
    OVMS_ServerNew(&srv); 
    OVMS_ServerSettingsSetLogLevel(serverSettings, OVMS_LOG_ERROR);  // Make the serving silent 
    OVMS_ModelsSettingsSetConfigPath(modelsSettings, "./config.json");  // Previously created file 
    OVMS_ServerStartFromConfigurationFile(srv, serverSettings, modelsSettings);  // Start the server 

‍

Input Data Preparation

Use OVMS_InferenceRequestInputSetData call, to provide input data with no additional copy operation. In InferenceRequestNew call, we can specify model name (the same as defined in config.json) and specific version (or 0 to use default). We also need to pass input names, data precision and shape information. In the example we provide 10 subsequent floating-point numbers, starting from 0.

    const char* MODEL_NAME = "dummy"; 
    const uint64_t MODEL_VERSION = 1; 
    const char* INPUT_NAME = "b"; 
    constexpr size_t NUM_OF_ELEMENTS = 10; 
    constexpr std::array SHAPE = {1, NUM_OF_ELEMENTS}; 
    OVMS_InferenceRequest* request; 
    OVMS_InferenceRequestNew(&request, srv, MODEL_NAME, MODEL_VERSION); 
    OVMS_InferenceRequestAddInput(request, INPUT_NAME, OVMS_DATATYPE_FP32, SHAPE.data(), SHAPE.size()); 
    std::array data{0, 1, 2, 3, 4, 5, 6, 7, 8, 9}; 
    OVMS_InferenceRequestInputSetData(request, INPUT_NAME, data.data(), sizeof(data), OVMS_BUFFERTYPE_CPU, 0); 

‍

Invoke Synchronous Inference

Simply call OVMS_Inference. This is required to pass response pointer and receive results in the next steps.

OVMS_InferenceResponse* response; 
    OVMS_Inference(srv, request, &response); 
  

‍

Read Results

Use call OVMS_InferenceResponseGetOutput API call to read the results. There are bunch of metadata we can read optionally, such as: precision, shape, buffer type and device ID. The expected output after addition should be:

1, 2, 3, 4, 5, 6, 7, 8, 9, 10,

const char* outputName; 
    OVMS_DataType dataType; 
    const uint64_t* shape; 
    uint32_t dimCount; 
    const void* outputData; 
    size_t byteSize; 
    OVMS_BufferType bufferType; 
    uint32_t deviceId; 
    OVMS_InferenceResponseGetOutput(response, 0, 
        &outputName, &dataType, &shape, &dimCount, &outputData, &byteSize, &bufferType, &deviceId); 
    for (int i = 0; i < NUM_OF_ELEMENTS; i++) 
        std::cout << ((float*)outputData)[i] << ", "; 
    std::cout << std::endl;

‍

Check the header file to learn more about the supported methods and their parameters.

‍

Compile and Run Application

In this example we omitted error handling and resource cleanup upon failure. Please refer to the full demo instructions for a more complete example.

‍

Performance Analysis

Using benchmarking tools from OpenVINO Runtime and both the C-API and gRPC API in OpenVINO Model Server, we can compare inference results via C-API to typical scenario of gRPC or direct integration of OpenVINO Runtime. The Resnet-50-tf model from Open Model Zoo was used for the testing below.

‍

Conclusion

With the new method of embedding OVMS into C++ applications, users can decrease inference latency even further by entirely skipping the networking part of model serving. The C-API is still in preview and has some limitations, but in its current state is ready to integrate into C++ applications. If you have questions or feedback, please file an issue on GitHub.

Read more:

• Complete API description: https://docs.openvino.ai/latest/ovms_docs_c_api.html  
• End to end demo: https://docs.openvino.ai/latest/ovms_demo_capi_inference_demo.html  

‍

Introduction

CPU plugin of OpenVINO™ toolkit as one of the most important part, which is powered by oneAPI Deep Neural Network Library (oneDNN) can help user achieve high performance inference of neural networks on Intel®x86-64 CPUs. The CPU plugin detects the Instruction Set Architecture (ISA) in the runtime and uses Just-in-Time (JIT) code generation to deploy the implementation optimized for the latest supported ISA.

In this blog, you will learn how layer primitives been optimized by implementation of ISA extensions and how to change the ISA extensions’ optimized kernel function at runtime for performance tuning and debugging.

After reading this blog, you will start to be proficient in AI workloads performance tuning and OpenVINO™ profiling on Intel® CPU architecture.

CPU Profiling

OpenVINO™ provide Application Program Interface (API) which is easy to turn on CPU profiling and analyze performance of each layer from the bottom level by executed kernel function. Firstly, enable performance counter profiling with executed device during device property configuration before model compiling with device. Learn detailed information from document of OpenVINO™ Configuring Devices.

core.set_property("CPU", ov::enable_profiling(true));

Then, you are allowed to get object of profiling info from inference requests which complied with the CPU device plugin.

auto perfCounts = infer_request.get_profiling_info();

Please note that performance profiling information generally can get after model inference. Refer below code implementation and add this part after model inference. You are possible to get status and performance of layer execution. Follow below code implement, you will get performance counter printing in order of the execution time from largest to smallest.

bool sort_pc_descend(const ov::ProfilingInfo& profiling1, const ov::ProfilingInfo& profiling2) {
    return profiling1.real_time > profiling2.real_time;
}
...
int tmain(int argc, tchar* argv[]) {
	...
	//objects init
	int layersize = 0;
	bool sort_pc_descend = 1;
	std::chrono::microseconds total = std::chrono::microseconds::zero();
	std::chrono::microseconds total_cpu = std::chrono::microseconds::zero();
	static const char* status_names[] = {"NOT_RUN", "OPTIMIZED_OUT", "EXECUTED"};

	//print row of title
	std::cout << "layerName\t"
	    << "execStatus\t"
	    << "layerType:"
	    << "execType\t";
	std::cout << "realTime (ms)\t"
	    << "cpuTime (ms)\t"
	    << " proportion(%)" << std::endl;

	//calculate executed layers total latency
	for (const auto& layer : perfCounts) {
	    if (std::string(status_names[(int)layer.status]).compare("EXECUTED") == 0) {
	        total += layer.real_time;
	        total_cpu += layer.cpu_time;
	    }
	}

	//print executed layer name, status, execution kernel funtion, CPU execution time and percentage of total model latency
	std::sort(perfCounts.begin(), perfCounts.end(), sort_pc_descend);
	for (const auto& layer : perfCounts) {
	    if (std::string(status_names[(int)layer.status]).compare("EXECUTED") == 0) {
	        std::cout << layer.node_name << "\t";
	        std::cout << ((int)layer.status < (sizeof(status_names) / sizeof(status_names[0]))
	                    ? status_names[(int)layer.status]
	                    : "INVALID_STATUS") << "\t";
	        std::cout << layer.node_type << ":" << layer.exec_type << "\t";
	        std::cout << std::to_string(layer.real_time.count() / 1000.0) << "\t"
	            << std::to_string(layer.cpu_time.count() / 1000.0) << "\t";
	        std::cout << (layer.real_time * 1.0 / total) * 100 << std::endl;
	        layersize += 1;
	    }
	}
}

CPU Dispatching

By enabling device profiling and printing exec_type of layers, you will get the specific kernel functions which powered by oneDNN during runtime execution. Use TensorFlow* ResNet 50 INT8 model for execution and pick the first 10 hotspot layers on 4th Gen Intel® Xeon Scalable processor (code named Sapphire Rapids) as an example:

From execution type of layers, it would be helpful to check which oneDNN kernel function used, and the actual precision of layer execution and the optimization from supported ISA on this platform.

Normally, oneDNN is able to detect to certain ISA, and OpenVINO™ allow to use latest ISA with higher priority. If you want to compare optimization rate between different ISA, can use the ONEDNN_MAX_CPU_ISA environment variable to limit processor features with older instruction sets. Follow this link to check oneDNN supported ISA.

Please note, Intel® Advanced Matrix Extensions (Intel® AMX) ISA start to be supported since 4th Gen Intel® Xeon Scalable processor. You can refer Intel® Product Specifications to check the supported instruction set of your current platform.

The ISAs are partially ordered:

·      SSE41 < AVX < AVX2 < AVX2_VNNI <AVX2_VNNI_2,

·      AVX2 < AVX512_CORE < AVX512_CORE_VNNI< AVX512_CORE_BF16 < AVX512_CORE_FP16 < AVX512_CORE_AMX <AVX512_CORE_AMX_FP16,

·      AVX2_VNNI < AVX512_CORE_FP16.

To use CPU dispatcher control, just set the value of ONEDNN_MAX_CPU_ISA environment variable before executable program which contains the OpenVINO™ device profiling printing, you can use benchmark_app as an example:

ONEDNN_MAX_CPU_ISA=AVX512_CORE_VNNI benchmark_app -m ~/models/public/resnet-50-tf/FP32-INT8/resnet-50-tf.xml -d CPU -pcsort sort

The benchmark_app provides the option which named “-pcsort” can report performance counters and order analysis information by order of layers execution time when set value of the option by “sort”.

In this case, we use above code implementation can achieve similar functionality of benchmark_app “-pcsort” option. User can consider try to add the code implementation into your own OpenVINO™ program like below:

ONEDNN_MAX_CPU_ISA=AVX512_CORE_VNNI ./classification_profiling ~/models/public/resnet-50-tf/FP32-INT8/resnet-50-tf.xml ./sample_640×426.bmp CPU

After setting the CPU dispatcher, the kernel execution function has been switched from AVX512_CORE_AMX to AVX512_CORE_VNNI. Then, the performance counters information would be like below:

You can easily find the hotspot layers of the same model would be changed when executed by difference kernel function which optimized by implementation of different ISA extensions. That is also the optimization differences between architecture platforms.

Tuning Tips

Users can refer the CPU dispatcher control and OpenVINO™ device profiling API to realize performance tuning of your inference program between CPU architectures. It will also be helpful to developer finding out the place where has the potential space of performance improvement.

For example, the hotspot layer generally should be compute-intensive operations like matrix-matrix multiplication; General vector operations which is not target to artificial intelligence (AI) / machine learning (ML) workloads cannot be optimized by Intel® AMX and Intel® Deep Learning Boost (Intel® DL Boost), and the memory accessing operations, like Transpose which maybe cannot parallelly optimized with instruction sets. If your inference model remains large memory accessing operations rather than compute-intensive operations, you probably need to be focusing on RAM bandwidth optimization.

‍

OpenVINO empowers developers to write deep learning application code once and deploy it on a wide range of Intel hardware with best-in-class performance. Previously, significant effort had to be spent configuring inference pipelines to squeeze optimal performance out of target hardware, and the effort had to be repeated whenever the application was ported to a new platform. The new Auto Device Plugin (AUTO) and automatic configuration features in OpenVINO make it easier for developers to unlock performance on multiple hardware targets without needing to spend time optimizing their application pipeline.

When an OpenVINO application is deployed in a system, the Auto Device Plugin automatically selects the best hardware target to inference the model with. OpenVINO then automatically configures the application to use optimal pipeline parameters based on the hardware capabilities and model size. Developers no longer need to write code for detecting hardware devices and explicitly configuring batch and stream parameters. High-level configuration is provided through performance hints that allow a developer to prioritize their application for either high throughput or minimal latency. AUTO and automatic device configuration make applications hardware-agnostic, allowing them to easily be ported to new hardware without any code changes.

The diagram in Figure 1 shows how OpenVINO’s features automatically configure an application for optimal performance, regardless of the target hardware. When the deep learning model is loaded, AUTO creates a transparent plugin interface to the available processor devices and automatically selects the most suitable device. OpenVINO configures the batch size and number of processing streams based on the selected hardware target, and the Auto-Batching feature automatically groups incoming data into optimally sized batches. AUTO and automatic configuration operate independently from each other, so developers can use either or both in their application.

AUTO and automatic configuration are available starting in the 2022.1 release of OpenVINO Runtime. To use these features, simply install OpenVINO Runtime on the target hardware. The API uses AUTO by default if no processor device is specified when loading a model. Set a “throughput” or “latency” performance hint when loading the model, and the API automatically configures the inference pipeline. Read on to learn more about AUTO, automatic configuration, performance hints, and how to use them in your application.

‍

Automatic Device Selection

Auto Device Plugin (AUTO) is a “virtual” device that provides a transparent interface to physical devices in the system. When an application is initialized, AUTO discovers the available processors and accelerators in the system (CPUs, integrated GPUs, discrete GPUs, VPUs) and selects the best device, based on a default device priority list or an optional user-provided priority list. It creates an interface between the application and device that executes inference requests in an optimized fashion. It enables an application to always achieve optimal performance in a system without the developer having to know beforehand what devices are available in the system.

Key Features and Benefits

Simple and flexible application deployment

Previously, developers needed to know details about target hardware and configure their application specifically for each device. AUTO removes the need to write dedicated code for specific devices. This enables an application to be written once and deployed to any supported hardware. It also allows the application to run on newer generations of hardware as they are released: the developer only needs to compile the application with the latest version of OpenVINO to run it on new hardware. This provides an instant increase in performance with little development time.

‍

Configurability

AUTO provides a configuration interface that is easy to use at a high level while still providing flexibility. Developers can simply specify “AUTO” as the device to tell the application to select the best device for the given model. They can also control which device is selected by providing a device candidate list and setting priorities for each device. 

Developers can also use performance hints to configure their application for latency or throughput. When the performance hint is throughput, OpenVINO will create more streams for parallel inferencing to achieve maximum processing bandwidth. In latency mode, OpenVINO creates fewer streams to utilize as many resources as possible to complete each inference quickly. Performance hints also help determine the optimal batch size for inferencing; this is discussed further in the “Performance Hints” section of this document.

‍

Improved first-inference latency

In applications that use accelerated processors like GPUs or VPUs, the time to first inference may be higher than average because it takes time to compile and load the deep learning model into the accelerator. AUTO solves this problem by starting the first inference with the CPU, which has minimal latency and no delays. As the first inference is being performed, AUTO continues to compile and load the model for the selected accelerator device, and then transparently switches over to that device when it is ready. This significantly reduces time to first inference, and is beneficial for applications that require immediate inference results on startup.

‍

How Automatic Device Selection Works

To choose the best device for inference, AUTO discovers which hardware targets are available in the system and matches the model to the best supported device, using the following process:

‍

1. AUTO discovers which devices are available using the Query Device API. The query reads an internal file that lists installed hardware plugins, confirms the hardware modules are present by communicating with them through drivers, and returns a list of available devices in the system.
   
   
2. AUTO checks the precision of the input model by reading the model file.
   
   
3. AUTO selects the best available device in the device priority table (shown in Table 1 below) that is capable of supporting the model’s precision.
   
   
4. AUTO attempts to compile the model on the selected device. If the model doesn’t compile (for example, if the device doesn’t support all the operations required by the model), AUTO tries to compile it on the next best device until compilation is successful. The CPU is the final fallback device, as it supports all operations and precisions.

‍

By default, AUTO uses the device priority list shown in Table 1. Developers can customize the table to provide their own device priority list and limit the devices that are available to run inferencing. AUTO will not try to run inference on devices that are not provided in the device list.

‍

Table 1. Default AUTO Device Priority List

‍

As mentioned, AUTO reduces the first inference latency by compiling and loading the model to the CPU first. As the model is loaded to the CPU and first inference is performed, AUTO steps through the rest of the process for selecting the device and compiling the model to that device. This way, devices that require a long time for model compilation do not impede inference as the application is being initialized.

AUTO also provides a model priority feature that enables developers to control which models are loaded to which devices when there are multiple models running on a system with multiple devices. Developers can set “MODEL_PRIORITY” as “HIGH”, “MEDIUM”, or “LOW” to configure which models should be allocated to the best resource. This allows developers to ensure models that are critical for an application are always loaded to the fastest device for processing, while less critical models are loaded to slower devices.

For example, consider a medical imaging application with models for segmenting and/or classifying injuries in X-ray images running on a system that has both a GPU and a CPU. The segmentation model is set to HIGH priority because it takes more processing power to inference, while the classification model is set to MEDIUM priority. If both models are loaded at the same time, the segmentation model will be loaded to the GPU (the higher priority device) and the classification model will be loaded to the CPU (the lower priority device). If only the classification model is loaded, it will be loaded to the GPU since the GPU isn’t occupied by the higher-priority model.

Automatic Device Configuration

The performance of a deep learning application can be improved by configuring runtime parameters to fully utilize the target hardware. There are several factors to take into consideration when optimizing inference for a certain device, such as batch size and number of streams. (See Runtime Inference Optimizations in OpenVINO documentation for more information.) The optimal configuration for these parameters depends on the architecture and memory of the target hardware, and they need to be re-determined when porting an application from one device to another. 

OpenVINO provides features that automatically configure an application to use optimal runtime parameters to achieve the best performance on any supported hardware target. These features are enabled through performance hints, which allow a user to specify whether their application should be optimized for latency or throughput. The automatic configuration eliminates the time and effort required to determine optimal configurations. It makes it simple to port to new devices or write one application to work on multiple devices. OpenVINO’s automatic configuration features currently work with CPU and GPU devices, and support for VPUs will be added in a future release.

‍

Performance Hints

OpenVINO allows users to provide high-level "performance hints" for setting latency-focused or throughput-focused inference modes. These performance hints are “latency” and “throughput.” The hints cause the runtime to automatically adjust runtime parameters, such as number of processing streams and inference batch size, to prioritize for reduced latency or high throughput. Performance hints are supported by CPU and GPU devices, and a future release of OpenVINO will add support for VPUs.

The performance hints do not require any device-specific settings and are portable between devices. Parameters are automatically configured based on whichever device is being used. This allows users to easily port applications between hardware targets without having to re-determine the best runtime parameters for the new device.

Latency performance hint

Latency is the amount of time it takes to process a single inference request and is usually measured in milliseconds (ms). In applications where data needs to be inferenced and acted on as quickly as possible (such as autonomous driving), low latency is desirable. When applications are run with the “latency” performance hint, OpenVINO determines the optimal number of parallel inference requests for minimizing latency while still maximizing the parallelization capabilities of the hardware. It automatically sets the number of processing streams to achieve the best latency.

To achieve the fastest latency, the processor device should process only one inference request at a time so all the compute resources are available for calculation. However, devices with multiple cores (such as multi-socket CPUs or multi-tile GPUs) can deliver multiple streams with the same latency as they would with a single stream. OpenVINO automatically checks the compute demands of the model, queries capabilities of the device, and selects the number of streams to be the minimum required to get the best latency. For CPUs, this is typically one stream for each socket. For GPUs, it’s typically one stream per tile.

Throughput performance hint

Throughput is the amount of data an inferencing pipeline can process at once, and it is usually measured in frames per second (FPS) or inferences per second. In applications where large amounts of data needs to be inferenced simultaneously (such as multi-camera video streams), high throughput is needed. To achieve high throughput, the runtime should focus on fully saturating the device with enough data to process. When applications are run with the “throughput” performance hint, OpenVINO maximizes the number of parallel inference requests to utilize all the threads available on the device. On GPU, it automatically sets the inference batch size to fill up the GPU memory available.

To configure the runtime for high throughput, OpenVINO automatically sets the number of streams to use based on the architecture of the device. For CPUs, it creates as many streams as there are cores available. For GPUs, it uses a combination of batch size and parallel streams to fully utilize the GPU’s memory and compute resources. To determine the optimal configuration on GPUs, OpenVINO will first check if the network supports batching. If it does, it loads the network with a batch size of one, determines how much memory is used for the single-batch network, and then scales the batch size and streams up to fill the entire GPU.

Batch size can also be explicitly specified in code when the model is loaded. This can be useful in applications where the number of incoming data sources is known and constant. For example, in an application that processes four camera streams, specify a batch size of four so that each set of frames from the cameras is processed in a single inference request. More information on batch configuration is given in the Auto-Batching section below.

‍

Auto-Batching

Auto-Batching is a new feature of OpenVINO that performs on-the-fly grouping of data inference requests in an application. As the application makes individual inference requests, Auto-Batching transparently collects them into a batch. When the batch is full (or when a timeout limit is reached), OpenVINO executes inference on the whole batch. In short, it takes care of batching data efficiently so the developer doesn’t have to worry about it.

The Auto-Batching feature is controlled by the configuration parameter “ALLOW_AUTO_BATCHING”, which is enabled by default. Auto-Batching is activated when all of the following are true:

• ALLOW_AUTO_BATCHING is true
• The model is loaded to the target device with the throughput performance hint
• The target device supports batching (such as GPU)
• The model topology supports batching

‍

When Auto-Batching is activated, OpenVINO automatically determines the optimal batch size for an application based on model size and hardware capabilities. Developers can also explicitly specify the batch size when loading the model. While the inference pipeline is active, individual inference requests are gathered into a batch and then executed when the batch is full.

Auto-Batching also has a timeout feature that is configurable by the developer. If there aren’t enough individual requests collected within the developer-specified time limit, batch execution will fall back to just using individual inference requests. For example, a developer may specify a timeout limit of 500 ms and a batch size of 16 for a video processing inference pipeline. Once 16 frames are gathered, a batch inference request is made. If only 13 frames arrive before the 500 ms timeout is hit, the application will perform individual inference requests on each of the 13 frames. While the timeout feature makes the pipeline robust to interruptions in incoming data, hitting the timeout limit heavily reduces the performance. To avoid this, developers should make sure there is enough incoming data to fill the batch within the time limit in typical conditions.

Auto-Batching, when combined with OpenVINO's automatic configuration features that determine optimal batch size and number of streams, provides a powerful benefit to the developer. The developer can utilize the full power of the target device with only using one line of code. Best of all, when an application is used on a different device, it will automatically reconfigure itself to achieve optimal performance with zero effort from the developer.

How to Use AUTO and Performance Hints

Using AUTO and automatic configuration with performance hints only requires one line of code. The functionality centers around the “ie.compile_model” method, which is used to compile a model and load it into device memory. The method accepts various configuration parameters that allow a user to provide high-level control over the pipeline.

Here are several Python examples showing how to configure a model and pipeline with the ie.compile_model method. The first example also shows how to import the OpenVINO Core model, initialize it, and read a model before calling ie.compile_model.

‍

Example 1. Load a model on AUTO device

from openvino.runtime import Core

ie = Core() # Initialize inference engine
model = ie.read_model(model="model.xml") # Read the model file

# Load model onto AUTO device
compiled_model = ie.compile_model(model=model, device_name="AUTO")

‍

Example 2. Load a model on AUTO device with performance hints

# Load model using the THROUGHPUT performance hint
compiled_model = core.compile_model(model=model, device_name="AUTO", config={"PERFORMANCE_HINT":"THROUGHPUT"}

# Alternatively, load model using the LATENCY performance hint
compiled_model = core.compile_model(model=model, device_name="AUTO", config={"PERFORMANCE_HINT":"LATENCY"})

‍

Example 3. Provide a list of device candidates which AUTO may use when loading a model

# Specify that AUTO can use either the GPU or CPU device
compiled_model = core.compile_model(model=model, device_name="AUTO:GPU,CPU")

‍

‍Example 4. Load multiple models with HIGH, MEDIUM, and LOW priorities

# Load three models and set them as HIGH, MEDIUM, and LOW priority
compiled_model0 = core.compile_model(model=model, device_name="AUTO", config={"MODEL_PRIORITY":"HIGH"})

compiled_model1 = core.compile_model(model=model, device_name="AUTO", config={"MODEL_PRIORITY":"MEDIUM"})

compiled_model2 = core.compile_model(model=model, device_name="AUTO", config={"MODEL_PRIORITY":"LOW"})

‍

‍Example 5. Load a model to GPU and use Auto-Batching with an explicitly set batch size

# Load model to GPU in throughput mode, with batch size set to 4
# (i.e. Auto-Batching collects 4 individual batches and then runs them all at once)
compiled_model = core.compile_model(model, "BATCH:GPU(4)", {"PERFORMANCE_HINT": "THROUGHPUT"})

‍

For a more in-depth example of how to use AUTO and automatic configuration, please visit the Automatic Device Selection with OpenVINO Jupyter notebook in the OpenVINO notebooks repository. It provides an end-to-end example that shows:

1. How to download a model from Open Model Zoo and convert it to OpenVINO IR format with Model Optimizer
2. How to load a model to AUTO device
3. The improvement in first inference latency when using AUTO device
4. How to perform asynchronous inferencing on data batches in throughput or latency mode
5. A performance comparison between throughput and latency modes

The OpenVINO Benchmark App also serves as a useful tool for experimenting with devices and batching to see how performance changes under various configurations. The Benchmark App supports automatic device selection and performance hints for throughput or latency.

Where to Learn More

To learn more please visit auto device plugin and automatic configuration pages in OpenVINO documentation. They provide more information about how to use and configure them in an application.

OpenVINO also provides an example notebook explaining how to use AUTO and showing how it improves performance. The notebook can be downloaded and run on a development machine where OpenVINO Developer Tools have been installed. Visit the notebook at this link: Automatic Device Selection with OpenVINO.

To learn more about OpenVINO toolkit and how to use it to build optimized deep learning applications, visit the Get Started page. OpenVINO also provides a number of example notebooks showing how to use it for basic applications like object detection and speech recognition on the Tutorials page.