# GNA Plugin¶

## Introducing the GNA Plugin¶

The Intel® Gaussian & Neural Accelerator is a low-power neural coprocessor for continuous inference at the edge.

Intel® GNA is not intended to replace typical inference devices such as the CPU, graphics processing unit (GPU), or vision processing unit (VPU). It is designed for offloading continuous inference workloads including but not limited to noise reduction or speech recognition to save power and free CPU resources.

The GNA plugin provides a way to run inference on Intel® GNA, as well as in the software execution mode on CPU.

## Devices with Intel® GNA¶

Devices with Intel® GNA support:

Note

On platforms where Intel® GNA is not enabled in the BIOS, the driver cannot be installed, so the GNA plugin uses the software emulation mode only.

## Intel® GNA Generational Differences¶

The first and second versions of Intel® GNA found in 10th and 11th generation Intel® Core™ Processors may be considered to be functionally equivalent. Intel® GNA 2.0 provided performance improvement with respect to Intel® GNA 1.0. Starting with 12th Generation Intel® Core™ Processors (formerly codenamed Alder Lake), support for Intel® GNA 3.0 features is being added.

In the rest of this documentation, “GNA 2.0” refers to Intel® GNA hardware delivered on 10th and 11th generation Intel® Core™ processors, and the term “GNA 3.0” will be used to refer to GNA hardware delivered on 12th generation Intel® Core™ processors.

Initially, a limited subset of Intel® GNA 3.0 features are added to the previous feature set including the following:

• 2D VALID Convolution With Small 2D Kernels: Two-dimensional convolutions with the following kernel dimensions [H,W] are supported: [1,1], [2,2], [3,3], [2,1], [3,1], [4,1], [5,1], [6,1], [7,1], [1,2], or [1,3]. Input tensor dimensions are limited to [1,8,16,16] <= [N,C,H,W] <= [1,120,384,240]. Up to 384 channels C may be used with a subset of kernel sizes (see table below). Up to 256 kernels (output channels) are supported. Pooling is limited to pool shapes of [1,1], [2,2], or [3,3]. Not all combinations of kernel shape and input tensor shape are supported (see the tables below for exact limitations).

The tables below show that the exact limitation on the input tensor width W depends on the number of input channels C (indicated as Ci below) and the kernel shape. There is much more freedom to choose the input tensor height and number of output channels.

## Initially Supported Subset of Intel® GNA 2D Convolutions¶

The following tables provide a more explicit representation of the Intel(R) GNA 3.0 2D convolution operations initially supported. The limits depend strongly on number of input tensor channels (Ci) and the input tensor width (W). Other factors are kernel height (KH), kernel width (KW), pool height (PH), pool width (PW), horizontal pool step (SH), and vertical pool step (PW). For example, the first table shows that for a 3x3 kernel with max pooling, only square pools are supported, and W is limited to 87 when there are 64 input channels.

Table of Maximum Input Tensor Widths (W) vs. Rest of Parameters (Input and Kernel Precision: 2 bytes)

KH

KW

PH

PW

SH

SW

H

W Ci=8 Co=256

W Ci=16 Co=256

W Ci=32 Co=256

W Ci=64 Co=256

W Ci=128 Co=256

W Ci=256 Co=256

W Ci=384 Co=256

1

1

1

1

1

1

128

240

240

240

240

240

240

170

1

1

1

1

1

1

256

240

240

240

240

240

128

85

1

1

1

1

1

1

384

240

240

240

240

170

85

56

1

2

1

1

1

1

128

240

240

240

240

1

2

1

1

1

1

256

240

240

240

240

1

2

1

1

1

1

384

240

240

240

240

1

3

1

1

1

1

128

240

240

240

240

1

3

1

1

1

1

256

240

240

240

240

1

3

1

1

1

1

384

240

240

240

240

2

1

1

1

1

1

128

192

192

192

192

192

192

128

2

1

1

1

1

1

256

192

192

192

192

192

128

85

2

1

1

1

1

1

384

192

192

192

192

170

85

56

2

2

1

1

1

1

128

193

193

193

193

2

2

1

1

1

1

256

193

193

193

193

2

2

1

1

1

1

384

193

193

193

193

2

2

2

2

1

1

128

193

193

192

179

2

2

2

2

1

1

256

193

193

192

179

2

2

2

2

1

1

384

193

193

192

179

2

2

2

2

1

2

128

193

193

192

179

2

2

2

2

1

2

256

193

193

192

179

2

2

2

2

1

2

384

193

193

192

179

2

2

2

2

2

1

128

193

193

192

179

2

2

2

2

2

1

256

193

193

192

179

2

2

2

2

2

1

384

193

193

192

179

2

2

2

2

2

2

128

193

193

192

179

2

2

2

2

2

2

256

193

193

192

179

2

2

2

2

2

2

384

193

193

192

179

3

1

1

1

1

1

128

128

128

128

128

128

85

42

3

1

1

1

1

1

256

128

128

128

128

128

85

42

3

1

1

1

1

1

384

128

128

128

128

128

85

42

3

3

1

1

1

1

128

130

130

130

87

3

3

1

1

1

1

256

130

130

130

87

3

3

1

1

1

1

384

130

130

130

87

3

3

2

2

1

1

128

130

130

126

87

3

3

2

2

1

1

256

130

130

126

87

3

3

2

2

1

1

384

130

130

126

87

3

3

2

2

1

2

128

130

130

126

87

3

3

2

2

1

2

256

130

130

126

87

3

3

2

2

1

2

384

130

130

126

87

3

3

2

2

2

1

128

130

130

126

87

3

3

2

2

2

1

256

130

130

126

87

3

3

2

2

2

1

384

130

130

126

87

3

3

2

2

2

2

128

130

130

126

87

3

3

2

2

2

2

256

130

130

126

87

3

3

2

2

2

2

384

130

130

126

87

3

3

3

3

1

1

128

130

128

118

87

3

3

3

3

1

1

256

130

128

118

87

3

3

3

3

1

1

384

130

128

118

87

3

3

3

3

1

2

128

130

128

118

87

3

3

3

3

1

2

256

130

128

118

87

3

3

3

3

1

2

384

130

128

118

87

3

3

3

3

1

3

128

130

128

118

87

3

3

3

3

1

3

256

130

128

118

87

3

3

3

3

1

3

384

130

128

118

87

3

3

3

3

2

1

128

130

128

118

87

3

3

3

3

2

1

256

130

128

118

87

3

3

3

3

2

1

384

130

128

118

87

3

3

3

3

2

2

128

130

128

118

87

3

3

3

3

2

2

256

130

128

118

87

3

3

3

3

2

2

384

130

128

118

87

3

3

3

3

2

3

128

130

128

118

87

3

3

3

3

2

3

256

130

128

118

87

3

3

3

3

2

3

384

130

128

118

87

3

3

3

3

3

1

128

130

128

118

87

3

3

3

3

3

1

256

130

128

118

87

3

3

3

3

3

1

384

130

128

118

87

3

3

3

3

3

2

128

130

128

118

87

3

3

3

3

3

2

256

130

128

118

87

3

3

3

3

3

2

384

130

128

118

87

3

3

3

3

3

3

128

130

128

118

87

3

3

3

3

3

3

256

130

128

118

87

3

3

3

3

3

3

384

130

128

118

87

4

1

1

1

1

1

128

96

96

96

96

96

64

32

4

1

1

1

1

1

256

96

96

96

96

96

64

32

4

1

1

1

1

1

384

96

96

96

96

96

64

32

5

1

1

1

1

1

128

76

76

76

76

51

25

5

1

1

1

1

1

256

76

76

76

76

51

25

5

1

1

1

1

1

384

76

76

76

76

51

25

6

1

1

1

1

1

128

64

64

64

64

42

21

6

1

1

1

1

1

256

64

64

64

64

42

21

6

1

1

1

1

1

384

64

64

64

64

42

21

7

1

1

1

1

1

128

54

54

54

54

36

7

1

1

1

1

1

256

54

54

54

54

36

7

1

1

1

1

1

384

54

54

54

54

36

Table of Maximum Input Tensor Widths (W) vs. Rest of Parameters (Input and Kernel Precision: 1 bytes)

KH

KW

PH

PW

SH

SW

H

W Ci=8 Co=256

W Ci=16 Co=256

W Ci=32 Co=256

W Ci=64 Co=256

W Ci=128 Co=256

W Ci=256 Co=256

W Ci=384 Co=256

1

1

1

1

1

1

128

240

240

240

240

240

240

240

1

1

1

1

1

1

256

240

240

240

240

240

240

170

1

1

1

1

1

1

384

240

240

240

240

240

170

113

1

2

1

1

1

1

128

240

240

240

240

240

240

240

1

2

1

1

1

1

256

240

240

240

240

240

240

170

1

2

1

1

1

1

384

240

240

240

240

240

170

113

1

3

1

1

1

1

128

240

240

240

240

240

1

3

1

1

1

1

256

240

240

240

240

240

1

3

1

1

1

1

384

240

240

240

240

240

2

1

1

1

1

1

128

192

192

192

192

192

192

192

2

1

1

1

1

1

256

192

192

192

192

192

192

170

2

1

1

1

1

1

384

192

192

192

192

192

170

113

2

2

1

1

1

1

128

193

193

193

193

193

193

129

2

2

1

1

1

1

256

193

193

193

193

193

193

129

2

2

1

1

1

1

384

193

193

193

193

193

170

113

3

1

1

1

1

1

128

128

128

128

128

128

128

85

3

1

1

1

1

1

256

128

128

128

128

128

128

85

3

1

1

1

1

1

384

128

128

128

128

128

128

85

3

3

1

1

1

1

128

130

130

130

130

87

3

3

1

1

1

1

256

130

130

130

130

87

3

3

1

1

1

1

384

130

130

130

130

87

4

1

1

1

1

1

128

96

96

96

96

96

96

64

4

1

1

1

1

1

256

96

96

96

96

96

96

64

4

1

1

1

1

1

384

96

96

96

96

96

96

64

5

1

1

1

1

1

128

76

76

76

76

76

51

51

5

1

1

1

1

1

256

76

76

76

76

76

51

51

5

1

1

1

1

1

384

76

76

76

76

76

51

51

6

1

1

1

1

1

128

64

64

64

64

64

42

21

6

1

1

1

1

1

256

64

64

64

64

64

42

21

6

1

1

1

1

1

384

64

64

64

64

64

42

21

7

1

1

1

1

1

128

54

54

54

54

54

36

18

7

1

1

1

1

1

256

54

54

54

54

54

36

18

7

1

1

1

1

1

384

54

54

54

54

54

36

18

Note

The above limitations only apply to the new hardware 2D convolution operation. When possible, the Intel® GNA plugin graph compiler flattens 2D convolutions so that the second generation Intel® GNA 1D convolution operations (without these limitations) may be used. The plugin will also flatten 2D convolutions regardless of the sizes if GNA 2.0 compilation target is selected (see below).

## Intel® GNA Forward and Backward Compatibility¶

In the general case, there is no guarantee that a model compiled for GNA 2.0 will run on GNA 3.0, or vice versa.

However, in most cases, networks compiled for GNA 2.0 will run as expected on GNA 3.0, although the performance may be worse compared to the case when a network is compiled specifically for the latter. The exception is networks with convolutions with the number of filters greater than 8192 (see the Models and Layers Limitations section).

Networks compiled for GNA 3.0 should run on GNA 2.0 with incompatible layers emulated on CPU.

You can use the following options KEY_GNA_EXEC_TARGET and KEY_GNA_COMPILE_TARGET options to check interoperability (see the Supported Configuration Parameters section below):

KEY_GNA_EXEC_TARGET, KEY_GNA_COMPILE_TARGET

GNA_EXEC_TARGET, GNA_COMPILE_TARGET

## Drivers and Dependencies¶

Intel® GNA hardware requires a driver to be installed on the system.

• Windows* OS: Intel® GNA driver for Windows is available through Windows Update*

## Models and Layers Limitations¶

Because of specifics of hardware architecture, Intel® GNA supports a limited set of layers, their kinds and combinations. For example, you should not expect the GNA Plugin to be able to run computer vision models, except those specifically adapted for the GNA Plugin, because the plugin does not fully support 2D convolutions.

For the list of supported layers, see the GNA column of the Supported Layers section in Supported Devices.

Limitations include:

• Only 1D convolutions are natively supported.

• The number of output channels for convolutions must be a multiple of 4.

• The maximum number of filters is 65532 for GNA 2.0 and 8192 for GNA 3.0.

• Permute layer support is limited to the cases where no data reordering is needed or when reordering is happening for two dimensions, at least one of which is not greater than 8.

• Splits and concatenations are supported for continuous portions of memory (e.g., split of 1,2,3,4 to 1,1,3,4 and 1,1,3,4 or concats of 1,2,3,4 and 1,2,3,5 to 2,2,3,4).

### Support for 2D Convolutions in Previous Generations of GNA Hardware¶

The Intel® GNA 1.0 and 2.0 hardware natively supports only 1D convolutions.

However, 2D convolutions can be mapped to 1D when a convolution kernel moves in a single direction. GNA Plugin performs such a transformation for Kaldi nnet1 convolution. From this perspective, the Intel® GNA hardware convolution operation accepts an NHWC input and produces an NHWC output. Because OpenVINO™ only supports the NCHW layout, you may need to insert Permute layers before or after convolutions.

For example, the Kaldi model optimizer inserts such a permute after convolution for the rm_cnn4a network. This Permute layer is automatically removed by the GNA Plugin, because the Intel® GNA hardware convolution layer already produces the required NHWC result.

## Operation Precision¶

Intel® GNA essentially operates in the low-precision mode, which represents a mix of 8-bit (I8), 16-bit (I16), and 32-bit (I32) integer computations. Outputs calculated using a reduced integer precision are different from the scores calculated using the floating point format, for example, FP32 outputs calculated on CPU using the Inference Engine CPU Plugin.

Unlike other plugins supporting low-precision execution, the GNA plugin can calculate quantization factors at the model loading time, so you can run a model without calibration using the Post-Training Optimization Tool. However, this mode may not provide satisfactory accuracy because the internal quantization algorithm is based on heuristics which may or may not be efficient, depending on the model and dynamic range of input data.

Starting with 2021.4 release of OpenVINO, GNA plugin users are encouraged to use the POT API Usage sample for GNA to get a model with quantization hints based on statistics for the provided dataset.

## Execution Modes¶

Mode

Description

KEY_GNA_AUTO

Uses Intel® GNA if available, otherwise uses software execution mode on CPU.

KEY_GNA_HW

Uses Intel® GNA if available, otherwise raises an error.

KEY_GNA_SW

Deprecated. Executes the GNA-compiled graph on CPU performing calculations in the same precision as the Intel® GNA, but not in the bit-exact mode.

KEY_GNA_SW_EXACT

Executes the GNA-compiled graph on CPU performing calculations in the same precision as the Intel® GNA in the bit-exact mode.

KEY_GNA_HW_WITH_SW_FBACK

Uses Intel® GNA if available, otherwise raises an error. If the hardware queue is not empty, automatically falls back to CPU in the bit-exact mode.

KEY_GNA_SW_FP32

Executes the GNA-compiled graph on CPU but substitutes parameters and calculations from low precision to floating point (FP32).

Mode

Description

GNA_AUTO

Uses Intel® GNA if available, otherwise uses software execution mode on CPU.

GNA_HW

Uses Intel® GNA if available, otherwise raises an error.

GNA_SW

Deprecated. Executes the GNA-compiled graph on CPU performing calculations in the same precision as the Intel® GNA, but not in the bit-exact mode.

GNA_SW_EXACT

Executes the GNA-compiled graph on CPU performing calculations in the same precision as the Intel® GNA in the bit-exact mode.

GNA_HW_WITH_SW_FBACK

Uses Intel® GNA if available, otherwise raises an error. If the hardware queue is not empty, automatically falls back to CPU in the bit-exact mode.

GNA_SW_FP32

Executes the GNA-compiled graph on CPU but substitutes parameters and calculations from low precision to floating point (FP32).

## Supported Configuration Parameters¶

The plugin supports the configuration parameters listed below. The parameter names correspond to their usage through API keys, such as GNAConfigParams::KEY_GNA_DEVICE_MODE or PluginConfigParams::KEY_PERF_COUNT in C++ and GNA_DEVICE_MODE or PERF_COUNT in Python.

Parameter Name

Values

Default Value

Description

KEY_GNA_EXEC_TARGET

TARGET_2_0, TARGET_3_0

see below

Defines the execution target.

KEY_GNA_COMPILE_TARGET

TARGET_2_0, TARGET_3_0

see below

Defines the compilation target.

KEY_GNA_COMPACT_MODE

YES, NO

NO

Enables I/O buffers reuse to save space. Makes debugging harder.

KEY_GNA_SCALE_FACTOR

FP32 number

1.0

Sets the scale factor to use for input quantization.

KEY_GNA_DEVICE_MODE

GNA_AUTO, GNA_HW, GNA_HW_WITH_SW_FBACK GNA_SW_EXACT, GNA_SW_FP32

GNA_AUTO

One of the modes described in Execution Modes.

KEY_GNA_FIRMWARE_MODEL_IMAGE

std::string

""

Sets the name for the embedded model binary dump file.

KEY_GNA_PRECISION

I16, I8

I16

Sets the preferred integer weight resolution for quantization (ignored for models produced using POT).

KEY_PERF_COUNT

YES, NO

NO

Turns on performance counters reporting.

KEY_GNA_LIB_N_THREADS

1-127 integer number

1

Sets the number of GNA accelerator library worker threads used for inference computation in software modes.

The parameters are passed as std::map<std::string, std::string> on InferenceEngine::Core::LoadNetwork or InferenceEngine::SetConfig.

Normally, you do not need to select the execution target (KEY_GNA_EXEC_TARGET) and compilation target (KEY_GNA_COMPILE_TARGET). The default value for the execution target corresponds to available hardware, or latest hardware version supported by the plugin (i.e., GNA 3.0) if there is no GNA HW in the system. The compilation target is the same as the execution target by default. However, you may want to change the targets, for example, if you want to check how a model compiled for one generation would behave on the other generation (using the software emulation mode), or if you are willing to export a model for a specific version of GNA HW.

You can change the KEY_GNA_DEVICE_MODE parameter at run time using InferenceEngine::ExecutableNetwork::SetConfig, which works for any value excluding GNA_SW_FP32. This enables you to switch the execution between software emulation mode and hardware execution mode after the model is loaded.

Parameter Name

Values

Default Value

Description

GNA_EXEC_TARGET

TARGET_2_0, TARGET_3_0

_see below_

Defines the execution target.

GNA_COMPILE_TARGET

TARGET_2_0, TARGET_3_0

_see below_

Defines the compilation target.

GNA_COMPACT_MODE

YES, NO

NO

Enables I/O buffers reuse to save space. Makes debugging harder.

GNA_SCALE_FACTOR

FP32 number

1.0

Sets the scale factor to use for input quantization.

KEY_GNA_DEVICE_MODE

GNA_AUTO, GNA_HW, GNA_HW_WITH_SW_FBACK GNA_SW_EXACT, GNA_SW_FP32

GNA_AUTO

One of the modes described in Execution Modes.

GNA_FIRMWARE_MODEL_IMAGE

string

""

Sets the name for the embedded model binary dump file.

GNA_PRECISION

I16, I8

I16

Sets the preferred integer weight resolution for quantization (ignored for models produced using POT).

PERF_COUNT

YES, NO

NO

Turns on performance counters reporting.

GNA_LIB_N_THREADS

1-127 integer number

1

Sets the number of GNA accelerator library worker threads used for inference computation in software modes.

The parameters are passed as strings to IECore.load_network.

Normally, you do not need to select the execution target (GNA_EXEC_TARGET) and compilation target (GNA_COMPILE_TARGET). The default value for the execution target corresponds to available hardware, or latest hardware version supported by the plugin (i.e., GNA 3.0) if there is no GNA HW in the system. The compilation target is the same as the execution target by default. However, you may want to change the targets, for example, if you want to check how a model compiled for one generation would behave on the other generation (using the SW emulation mode), or if you are willing to export a model for a specific version of GNA HW.

You can change the GNA_DEVICE_MODE parameter at run time by sending a configuration dict to the IECore.load_network call, which works for any value excluding GNA_SW_FP32. This enables you to switch the execution between software emulation mode and hardware execution mode after the model is loaded.

## How to Interpret Performance Counters¶

With the following methods, you can collect performance counters that provides various performance data about execution on GNA:

InferenceEngine::InferRequest::GetPerformanceCounts

The returned map stores a counter description as a key, and a counter value in the realTime_uSec field of the InferenceEngineProfileInfo structure.

openvino.inference_engine.InferRequest.get_perf_counts

The returned map stores a counter description as a key, and a counter value in the real_time field.

The current GNA implementation calculates counters for the whole utterance scoring and does not provide per-layer information. The API enables you to retrieve counter units in cycles, you can convert cycles to seconds as follows:

seconds = cycles / frequency

Refer to the table below to learn about the frequency of Intel® GNA inside a particular processor:

Processor

Frequency of Intel® GNA

Intel® Core™ processors

400MHz

Intel® processors formerly codenamed Elkhart Lake

200MHz

Performance counters provided for the time being:

• Scoring request performance results

• Number of total cycles spent on scoring in hardware including compute and memory stall cycles

• Number of stall cycles spent in hardware

## Multithreading Support in GNA Plugin¶

The GNA plugin supports the following configuration parameters for multithreading management:

KEY_GNA_LIB_N_THREADS

GNA_LIB_N_THREADS

By default, the GNA plugin uses one worker thread for inference computations. This parameter allows you to create up to 127 threads for software modes.

Note

Multithreading mode does not guarantee the same computation order as the order of issuing. Additionally, in this case, software modes do not implement any serializations.

## Network Batch Size¶

Intel® GNA plugin supports the processing of context-windowed speech frames in batches of 1-8 frames in one input blob using the following methods:

InferenceEngine::ICNNNetwork::setBatchSize

Increasing batch size only improves efficiency of Fully Connected layers.

Note

For networks with Convolutional, LSTM, or Memory layers, the only supported batch size is 1.

## Compatibility with Heterogeneous Plugin¶

Heterogeneous plugin was tested with the Intel® GNA as a primary device and CPU as a secondary device. To run inference of networks with layers unsupported by the GNA plugin, such as Softmax, use the Heterogeneous plugin with the HETERO:GNA,CPU configuration.

Note

Due to limitation of the Intel® GNA backend library, heterogenous support is limited to cases where in the resulted sliced graph, only one subgraph is scheduled to run on GNA_HW or GNA_SW devices.

## Recovery from Interruption by High-Priority Windows Audio Processes*¶

GNA is designed for real-time workloads such as noise reduction. For such workloads, processing should be time constrained, otherwise extra delays may cause undesired effects such as audio glitches. To make sure that processing can satisfy real-time requirements, the GNA driver provides a Quality of Service (QoS) mechanism, which interrupts requests that might cause high-priority Windows audio processes to miss the schedule, thereby causing long running GNA tasks to terminate early.

Applications should be prepared for this situation.

If an inference in the GNA_HW mode cannot be executed because of such an interruption, then the wait method returns the following status code:

InferRequest::Wait() returns status code StatusCode::INFER_NOT_STARTED.

InferRequest.wait returns status code INFER_NOT_STARTED.

In future releases, it will be changed to a more meaningful status code.

Any application working with GNA must properly react to this code. One of the strategies to adapt an application:

1. Immediately switch to the GNA_SW emulation mode:

std::map<std::string, Parameter> newConfig;
newConfig[GNAConfigParams::KEY_GNA_DEVICE_MODE] = Parameter("GNA_SW_EXACT");
executableNet.SetConfig(newConfig);

from openvino.inference_engine import IECore

ie = IECore()
new_cfg = {'GNA_DEVICE_MODE' : 'GNA_SW_EXACT'}

2. Resubmit and switch back to GNA_HW expecting that the competing application has finished.

Note

This method is deprecated since a new automatic QoS mode has been introduced in 2021.4.1 release of OpenVINO™ (see below).

## GNA3 Automatic QoS Feature on Windows*¶

Starting with 2021.4.1 release of OpenVINO and 03.00.00.1363 version of Windows* GNA driver, a new execution mode GNA_HW_WITH_SW_FBACK is introduced to assure that workloads satisfy real-time execution. In this mode, the GNA driver automatically falls back on CPU for a particular infer request if the HW queue is not empty, so there is no need for explicitly switching between GNA and CPU.

Note

Due to the “first come - first served” nature of GNA driver and the QoS feature, this mode may lead to increased CPU consumption

if there are several clients using GNA simultaneously. Even a lightweight competing infer request which has not been cleared at the time when the user’s GNA client process makes its request, can cause the user’s request to be executed on CPU, thereby unnecessarily increasing CPU utilization and power.