Supported Neural Network Operators

Neural network operators can be directly supported by the hardware, decomposed into hardware-compatible components, or approximated. While the hardware-supported operators only change with new chip generations, the coverage of decomposed and approximated operators increases with each new software release. This page lists the latest support for operators across various neural network frameworks.

keras

keras.activations.elu

Reference: keras.activations.elu

keras.activations.exponential

Reference: keras.activations.exponential

keras.activations.gelu

Reference: keras.activations.gelu

keras.activations.linear

Reference: keras.activations.linear

keras.activations.relu

Reference: keras.activations.relu

keras.activations.selu

Reference: keras.activations.selu

keras.activations.sigmoid

Reference: keras.activations.sigmoid

keras.activations.softmax

Reference: keras.activations.softmax

keras.activations.softplus

Reference: keras.activations.softplus

keras.activations.softsign

Reference: keras.activations.softsign

keras.activations.swish

Reference: keras.activations.swish

keras.activations.tanh

Reference: keras.activations.tanh

keras.layers.Activation

Reference: keras.layers.Activation

keras.layers.Add

Reference: keras.layers.Add

keras.layers.AlphaDropout

Reference: keras.layers.AlphaDropout

keras.layers.Attention

Reference: keras.layers.Attention

keras.layers.AveragePooling1D

Reference: keras.layers.AveragePooling1D

keras.layers.AveragePooling2D

Reference: keras.layers.AveragePooling2D

keras.layers.AveragePooling3D

Reference: keras.layers.AveragePooling3D

keras.layers.Average

Reference: keras.layers.Average

keras.layers.BatchNormalization

Reference: keras.layers.BatchNormalization

keras.layers.Concatenate

Reference: keras.layers.Concatenate

keras.layers.Conv1DTranspose

Reference: keras.layers.Conv1DTranspose

keras.layers.Conv1D

Reference: keras.layers.Conv1D

keras.layers.Conv2DTranspose

Reference: keras.layers.Conv2DTranspose

keras.layers.Conv2D

Reference: keras.layers.Conv2D

Notes:

• kernel_size <= input_size

keras.layers.Conv3DTranspose

Reference: keras.layers.Conv3DTranspose

keras.layers.Conv3D

Reference: keras.layers.Conv3D

keras.layers.Cropping1D

Reference: keras.layers.Cropping1D

keras.layers.Cropping2D

Reference: keras.layers.Cropping2D

keras.layers.Cropping3D

Reference: keras.layers.Cropping3D

keras.layers.Dense

Reference: keras.layers.Dense

keras.layers.DepthwiseConv1D

Reference: keras.layers.DepthwiseConv1D

keras.layers.DepthwiseConv2D

Reference: keras.layers.DepthwiseConv2D

keras.layers.Dot

Reference: keras.layers.Dot

keras.layers.Dropout

Reference: keras.layers.Dropout

keras.layers.ELU

Reference: keras.layers.ELU

keras.layers.Flatten

Reference: keras.layers.Flatten

keras.layers.GaussianDropout

Reference: keras.layers.GaussianDropout

keras.layers.GaussianNoise

Reference: keras.layers.GaussianNoise

keras.layers.GlobalAveragePooling1D

Reference: keras.layers.GlobalAveragePooling1D

keras.layers.GlobalAveragePooling2D

Reference: keras.layers.GlobalAveragePooling2D

keras.layers.GlobalAveragePooling3D

Reference: keras.layers.GlobalAveragePooling3D

keras.layers.GlobalMaxPooling1D

Reference: keras.layers.GlobalMaxPooling1D

keras.layers.GlobalMaxPooling2D

Reference: keras.layers.GlobalMaxPooling2D

keras.layers.GlobalMaxPooling3D

Reference: keras.layers.GlobalMaxPooling3D

keras.layers.InputLayer

Reference: keras.layers.InputLayer

keras.layers.LayerNormalization

Reference: keras.layers.LayerNormalization

keras.layers.LeakyReLU

Reference: keras.layers.LeakyReLU

keras.layers.Linear

Reference: keras.layers.Linear

keras.layers.MaxPooling1D

Reference: keras.layers.MaxPooling1D

keras.layers.MaxPooling2D

Reference: keras.layers.MaxPooling2D

keras.layers.MaxPooling3D

Reference: keras.layers.MaxPooling3D

keras.layers.Maximum

Reference: keras.layers.Maximum

keras.layers.Minimum

Reference: keras.layers.Minimum

keras.layers.MultiHeadAttention

Reference: keras.layers.MultiHeadAttention

keras.layers.Multiply

Reference: keras.layers.Multiply

keras.layers.Normalization

Reference: keras.layers.Normalization

keras.layers.PReLU

Reference: keras.layers.PReLU

keras.layers.Permute

Reference: keras.layers.Permute

keras.layers.ReLU

Reference: keras.layers.ReLU

keras.layers.RepeatVector

Reference: keras.layers.RepeatVector

keras.layers.Rescaling

Reference: keras.layers.Rescaling

keras.layers.Reshape

Reference: keras.layers.Reshape

keras.layers.SeparableConv1D

Reference: keras.layers.SeparableConv1D

keras.layers.SeparableConv2D

Reference: keras.layers.SeparableConv2D

keras.layers.Softmax

Reference: keras.layers.Softmax

keras.layers.SpatialDropout1D

Reference: keras.layers.SpatialDropout1D

keras.layers.SpatialDropout2D

Reference: keras.layers.SpatialDropout2D

keras.layers.SpatialDropout3D

Reference: keras.layers.SpatialDropout3D

keras.layers.Subtract

Reference: keras.layers.Subtract

keras.layers.TensorFlowOpLayer

Reference: keras.layers.TensorFlowOpLayer

keras.layers.ThresholdedReLU

Reference: keras.layers.ThresholdedReLU

keras.layers.UpSampling1D

Reference: keras.layers.UpSampling1D

keras.layers.UpSampling2D

Reference: keras.layers.UpSampling2D

keras.layers.UpSampling3D

Reference: keras.layers.UpSampling3D

keras.layers.ZeroPadding1D

Reference: keras.layers.ZeroPadding1D

keras.layers.ZeroPadding2D

Reference: keras.layers.ZeroPadding2D

keras.layers.ZeroPadding3D

Reference: keras.layers.ZeroPadding3D

onnx

Abs

Reference: onnx__Abs

Acos

Reference: onnx__Acos

Acosh

Reference: onnx__Acosh

Add

Reference: onnx__Add

And

Reference: onnx__And

Asin

Reference: onnx__Asin

Asinh

Reference: onnx__Asinh

Atan

Reference: onnx__Atan

Atanh

Reference: onnx__Atanh

AveragePool

Reference: onnx__AveragePool

BatchNormalization

Reference: onnx__BatchNormalization

Cast

Reference: onnx__Cast

Notes:

• The compiler will ignore the Cast operator, since all internal FMaps are BFloat16 .

Ceil

Reference: onnx__Ceil

Celu

Reference: onnx__Celu

Clip

Reference: onnx__Clip

Concat

Reference: onnx__Concat

ConvTranspose

Reference: onnx__ConvTranspose

Conv

Reference: onnx__Conv

Cos

Reference: onnx__Cos

Cosh

Reference: onnx__Cosh

DepthToSpace

Reference: onnx__DepthToSpace

Div

Reference: onnx__Div

Dropout

Reference: onnx__Dropout

Elu

Reference: onnx__Elu

Exp

Reference: onnx__Exp

Expand

Reference: onnx__Expand

FastGelu

Reference: onnx__FastGelu

Flatten

Reference: onnx__Flatten

Floor

Reference: onnx__Floor

Gelu

Reference: onnx__Gelu

Gemm

Reference: onnx__Gemm

GlobalAveragePool

Reference: onnx__GlobalAveragePool

GlobalMaxPool

Reference: onnx__GlobalMaxPool

GreaterOrEqual

Reference: onnx__GreaterOrEqual

Greater

Reference: onnx__Greater

HardSigmoid

Reference: onnx__HardSigmoid

HardSwish

Reference: onnx__HardSwish

Identity

Reference: onnx__Identity

InstanceNormalization

Reference: onnx__InstanceNormalization

LRN

Reference: onnx__LRN

Notes:

• Only support beta=0.75 (default) for now.

LayerNormalization

Reference: onnx__LayerNormalization

LayerNormalization

Reference: onnx__LayerNormalization

LeakyRelu

Reference: onnx__LeakyRelu

LessOrEqual

Reference: onnx__LessOrEqual

Less

Reference: onnx__Less

Log

Reference: onnx__Log

MatMul

Reference: onnx__MatMul

MaxPool

Reference: onnx__MaxPool

Max

Reference: onnx__Max

Mean

Reference: onnx__Mean

Min

Reference: onnx__Min

Mod

Reference: onnx__Mod

Mul

Reference: onnx__Mul

Neg

Reference: onnx__Neg

Not

Reference: onnx__Not

Or

Reference: onnx__Or

PRelu

Reference: onnx__PRelu

Pad

Reference: onnx__Pad

Pow

Reference: onnx__Pow

QLinearAdd

Reference: onnx__QLinearAdd

QLinearConv

Reference: onnx__QLinearConv

QuantizeLinear

Reference: onnx__QuantizeLinear

Reciprocal

Reference: onnx__Reciprocal

ReduceL1

Reference: onnx__ReduceL1

ReduceL2

Reference: onnx__ReduceL2

ReduceMax

Reference: onnx__ReduceMax

ReduceMean

Reference: onnx__ReduceMean

ReduceMin

Reference: onnx__ReduceMin

ReduceSum

Reference: onnx__ReduceSum

Relu

Reference: onnx__Relu

Reshape

Reference: onnx__Reshape

Resize

Reference: onnx__Resize

Round

Reference: onnx__Round

Selu

Reference: onnx__Selu

Shape

Reference: onnx__Shape

Notes:

• The compiler will statically evaluate and fold all shapes during

• compilation. We do not support variable shape inference.

Sigmoid

Reference: onnx__Sigmoid

Sign

Reference: onnx__Sign

Sin

Reference: onnx__Sin

Sinh

Reference: onnx__Sinh

Slice

Reference: onnx__Slice

Softmax

Reference: onnx__Softmax

Softplus

Reference: onnx__Softplus

Softsign

Reference: onnx__Softsign

SpaceToDepth

Reference: onnx__SpaceToDepth

Split

Reference: onnx__Split

Sqrt

Reference: onnx__Sqrt

Squeeze

Reference: onnx__Squeeze

Sub

Reference: onnx__Sub

Sum

Reference: onnx__Sum

Tan

Reference: onnx__Tan

Tanh

Reference: onnx__Tanh

ThresholdedRelu

Reference: onnx__ThresholdedRelu

Tile

Reference: onnx__Tile

Transpose

Reference: onnx__Transpose

Unsqueeze

Reference: onnx__Unsqueeze

Upsample

Reference: onnx__Upsample

Xor

Reference: onnx__Xor

tensorflow

Abs

Reference: Abs

AddN

Reference: AddN

AddV2

Reference: AddV2

Add

Reference: Add

AvgPool

Reference: AvgPool

BatchToSpaceND

Reference: BatchToSpaceND

BiasAdd

Reference: BiasAdd

Cast

Reference: Cast

ConcatV2

Reference: ConcatV2

Concat

Reference: Concat

Const

Reference: Const

Conv2DBackpropInput

Reference: Conv2DBackpropInput

Conv2D

Reference: Conv2D

Conv3D

Reference: Conv3D

DepthToSpace

Reference: DepthToSpace

DepthwiseConv2dNative

Reference: DepthwiseConv2dNative

Exp

Reference: Exp

ExpandDims

Reference: ExpandDims

Fill

Reference: Fill

FusedBatchNorm

Reference: FusedBatchNorm

GatherV2

Reference: GatherV2

IdentityN

Reference: IdentityN

Identity

Reference: Identity

LeakyRelu

Reference: LeakyRelu

MatMul

Reference: MatMul

MaxPool

Reference: MaxPool

Max

Reference: Max

Maximum

Reference: Maximum

Mean

Reference: Mean

Minimum

Reference: Minimum

MirrorPad

Reference: MirrorPad

Mul

Reference: Mul

Neg

Reference: Neg

NoOp

Reference: NoOp

Pack

Reference: Pack

Pad

Reference: Pad

Placeholder

Reference: Placeholder

Pow

Reference: Pow

Prod

Reference: Prod

RealDiv

Reference: RealDiv

Relu6

Reference: Relu6

Relu

Reference: Relu

Reshape

Reference: Reshape

ResizeBilinear

Reference: ResizeBilinear

ResizeNearestNeighbor

Reference: ResizeNearestNeighbor

Rsqrt

Reference: Rsqrt

Selu

Reference: Selu

Shape

Reference: Shape

Sigmoid

Reference: Sigmoid

Softmax

Reference: Softmax

Softsign

Reference: Softsign

SpaceToBatchND

Reference: SpaceToBatchND

Split

Reference: Split

Sqrt

Reference: Sqrt

Square

Reference: Square

SquaredDifference

Reference: SquaredDifference

Squeeze

Reference: Squeeze

StopGradient

Reference: StopGradient

StridedSlice

Reference: StridedSlice

Sub

Reference: Sub

Tanh

Reference: Tanh

Tile

Reference: Tile

Transpose

Reference: Transpose

tflite

abs

Reference: tfl.abs

add_n

Reference: tfl.add_n

add

Reference: tfl.add

average_pool_2d

Reference: tfl.average_pool_2d

batch_to_space_nd

Reference: tfl.batch_to_space_nd

Notes:

• This is only supported as a node in the subgraph of a decompsed dilated convolution

broadcast_args

Reference: tfl.broadcast_args

broadcast_to

Reference: tfl.broadcast_to

cast

Reference: tfl.cast

ceil

Reference: tfl.ceil

concatenation

Reference: tfl.concatenation

conv_2d

Reference: tfl.conv_2d

conv_3d

Reference: tfl.conv_3d

conv_3d_transpose

Reference: tfl.conv_3d_transpose

densify

Reference: tfl.densify

depth_to_space_nd

Reference: tfl.depth_to_space_nd

depthwise_conv_2d

Reference: tfl.depthwise_conv_2d

dequantize

Reference: tfl.dequantize

div

Reference: tfl.div

elu

Reference: tfl.elu

equal

Reference: tfl.equal

exp

Reference: tfl.exp

expand_dims

Reference: tfl.expand_dims

fill

Reference: tfl.fill

floor_div

Reference: tfl.floor_div

floor_mod

Reference: tfl.floor_mod

floor

Reference: tfl.floor

fully_connected

Reference: tfl.fully_connected

gelu

Reference: tfl.gelu

greater_equal

Reference: tfl.greater_equal

greater

Reference: tfl.greater

hard_swish

Reference: tfl.hard_swish

l2_normalization

Reference: tfl.l2_normalization

leaky_relu

Reference: tfl.leaky_relu

less_equal

Reference: tfl.less_equal

less

Reference: tfl.less

local_response_normalization

Reference: tfl.local_response_normalization

logical_and

Reference: tfl.logical_and

logical_not

Reference: tfl.logical_not

logical_or

Reference: tfl.logical_or

logistic

Reference: tfl.logistic

max_pool_2d

Reference: tfl.max_pool_2d

maximum

Reference: tfl.maximum

mean

Reference: tfl.mean

minimum

Reference: tfl.minimum

mirror_pad

Reference: tfl.mirror_pad

mul

Reference: tfl.mul

neg

Reference: tfl.neg

not_equal

Reference: tfl.not_equal

pack

Reference: tfl.pack

pad

Reference: tfl.pad

padv2

Reference: tfl.padv2

pow

Reference: tfl.pow

prelu

Reference: tfl.prelu

quantize

Reference: tfl.quantize

reduce_any

Reference: tfl.reduce_any

reduce

Reference: tfl.reduce

reduce_min

Reference: tfl.reduce_min

relu6

Reference: tfl.relu6

relu_n1_to_1

Reference: tfl.relu_n1_to_1

relu

Reference: tfl.relu

reshape

Reference: tfl.reshape

resize_bilinear

Reference: tfl.resize_bilinear

resize_nearest_neighbor

Reference: tfl.resize_nearest_neighbor

round

Reference: tfl.round

rsqrt

Reference: tfl.rsqrt

shape

Reference: tfl.shape

sin

Reference: tfl.sin

slice

Reference: tfl.slice

softmax

Reference: tfl.softmax

space_to_batch_nd

Reference: tfl.space_to_batch_nd

space_to_depth

Reference: tfl.space_to_depth

split

Reference: tfl.split

sqrt

Reference: tfl.sqrt

square

Reference: tfl.square

squared_difference

Reference: tfl.squared_difference

squeeze

Reference: tfl.squeeze

strided_slice

Reference: tfl.strided_slice

sub

Reference: tfl.sub

sum

Reference: tfl.sum

tanh

Reference: tfl.tanh

tile

Reference: tfl.tile

transpose_conv

Reference: tfl.transpose_conv

transpose

Reference: tfl.transpose

unpack

Reference: tfl.unpack

pytorch

Note

PyTorch models can be supported by exporting them to ONNX (for more information see tutorial on exporting to ONNX). Direct support for Pytorch 2 will be available in a future release.