NPU Operator Support¶
The tables show operator support for aries-rb, regulus-ra, and regulus-rb. regulus-rb-usb has the same operator scope and device-specific conditions as regulus-rb. Framework operators may be transformed into Mobilint IR operations during parsing.
Legend: ✅ Supported, ❌ Unsupported on this NPU.
Even a listed operator may fail to compile when the model graph falls outside the common input contract (4-D NHWC, static batch size 1, non-boolean input) or supported data shapes and attributes. Adjacent operators and graph rewrites or removal also affect support. Verify by compiling the model. For CPU execution, see CPU Offloading.
ONNX¶
| Operator | aries-rb |
regulus-ra |
regulus-rb |
|---|---|---|---|
Add |
✅ | ✅ | ✅ |
AveragePool |
✅ | ✅ | ✅ |
BatchNormalization |
✅ | ✅ | ✅ |
Celu |
✅ | ✅ | ✅ |
Clip |
✅ | ✅ | ✅ |
Concat |
✅ | ✅ | ✅ |
Conv |
✅ | ✅ | ✅ |
ConvTranspose |
✅ | ✅ | ✅ |
Cos |
✅ | ❌ | ✅ |
DepthToSpace |
✅ | ✅ | ✅ |
Div |
✅ | ✅ | ✅ |
Elu |
✅ | ✅ | ✅ |
Erf |
✅ | ✅ | ✅ |
Exp |
✅ | ✅ | ✅ |
Flatten |
✅ | ✅ | ✅ |
Gelu |
✅ | ✅ | ✅ |
Gemm |
✅ | ✅ | ✅ |
GroupNormalization |
✅ | ✅ | ✅ |
GRU |
✅ | ✅ | ✅ |
HardSigmoid |
✅ | ✅ | ✅ |
HardSwish |
✅ | ✅ | ✅ |
InstanceNormalization |
✅ | ✅ | ✅ |
LayerNormalization |
✅ | ✅ | ✅ |
LeakyRelu |
✅ | ✅ | ✅ |
Log |
✅ | ✅ | ✅ |
LSTM |
✅ | ✅ | ✅ |
MatMul |
✅ | ✅ | ✅ |
Max |
✅ | ✅ | ✅ |
MaxPool |
✅ | ✅ | ✅ |
Mul |
✅ | ✅ | ✅ |
Pow |
✅ | ✅ | ✅ |
PRelu |
✅ | ✅ | ✅ |
QuickGelu |
✅ | ✅ | ✅ |
Reciprocal |
✅ | ✅ | ✅ |
ReduceMax |
✅ | ✅ | ✅ |
ReduceMean |
✅ | ✅ | ✅ |
ReduceMin |
✅ | ✅ | ✅ |
ReduceSum |
✅ | ✅ | ✅ |
Relu |
✅ | ✅ | ✅ |
Reshape |
✅ | ✅ | ✅ |
Resize |
✅ | ✅ | ✅ |
RNN |
✅ | ✅ | ✅ |
Sigmoid |
✅ | ✅ | ✅ |
Sin |
✅ | ❌ | ✅ |
Slice |
✅ | ✅ | ✅ |
Softmax |
✅ | ❌ | ✅ |
Softplus |
✅ | ✅ | ✅ |
SpaceToDepth |
✅ | ✅ | ✅ |
Split |
✅ | ✅ | ✅ |
Sqrt |
✅ | ✅ | ✅ |
Squeeze |
✅ | ✅ | ✅ |
Sub |
✅ | ✅ | ✅ |
swish |
✅ | ✅ | ✅ |
Tanh |
✅ | ✅ | ✅ |
Transpose |
✅ | ✅ | ✅ |
Unsqueeze |
✅ | ✅ | ✅ |
Upsample |
✅ | ✅ | ✅ |
TensorFlow¶
| Operator | aries-rb |
regulus-ra |
regulus-rb |
|---|---|---|---|
Add |
✅ | ✅ | ✅ |
AddN |
✅ | ✅ | ✅ |
AddV2 |
✅ | ✅ | ✅ |
AvgPool |
✅ | ✅ | ✅ |
BatchMatMul |
✅ | ✅ | ✅ |
BatchMatMulV2 |
✅ | ✅ | ✅ |
BatchMatMulV3 |
✅ | ✅ | ✅ |
BiasAdd |
✅ | ✅ | ✅ |
ClipByValue |
✅ | ✅ | ✅ |
Concat |
✅ | ✅ | ✅ |
ConcatV2 |
✅ | ✅ | ✅ |
Conv2D |
✅ | ✅ | ✅ |
Conv2DBackpropInput |
✅ | ✅ | ✅ |
Cos |
✅ | ❌ | ✅ |
DepthwiseConv2dNative |
✅ | ✅ | ✅ |
Div |
✅ | ✅ | ✅ |
Elu |
✅ | ✅ | ✅ |
Erf |
✅ | ✅ | ✅ |
Erfc |
✅ | ✅ | ✅ |
Exp |
✅ | ✅ | ✅ |
ExpandDims |
✅ | ✅ | ✅ |
FusedBatchNorm |
✅ | ✅ | ✅ |
FusedBatchNormV2 |
✅ | ✅ | ✅ |
FusedBatchNormV3 |
✅ | ✅ | ✅ |
Inv |
✅ | ✅ | ✅ |
LeakyRelu |
✅ | ✅ | ✅ |
Log |
✅ | ✅ | ✅ |
MatMul |
✅ | ✅ | ✅ |
Max |
✅ | ✅ | ✅ |
Maximum |
✅ | ✅ | ✅ |
MaxPool |
✅ | ✅ | ✅ |
Mean |
✅ | ✅ | ✅ |
Min |
✅ | ✅ | ✅ |
Minimum |
✅ | ✅ | ✅ |
Mul |
✅ | ✅ | ✅ |
Pow |
✅ | ✅ | ✅ |
RealDiv |
✅ | ✅ | ✅ |
Reciprocal |
✅ | ✅ | ✅ |
Relu |
✅ | ✅ | ✅ |
Relu6 |
✅ | ✅ | ✅ |
Reshape |
✅ | ✅ | ✅ |
ResizeBilinear |
✅ | ✅ | ✅ |
ResizeNearestNeighbor |
✅ | ✅ | ✅ |
Sigmoid |
✅ | ✅ | ✅ |
Sin |
✅ | ❌ | ✅ |
Slice |
✅ | ✅ | ✅ |
Softmax |
✅ | ❌ | ✅ |
Softplus |
✅ | ✅ | ✅ |
Split |
✅ | ✅ | ✅ |
SplitV |
✅ | ✅ | ✅ |
Sqrt |
✅ | ✅ | ✅ |
Square |
✅ | ✅ | ✅ |
SquaredDifference |
✅ | ✅ | ✅ |
Squeeze |
✅ | ✅ | ✅ |
StridedSlice |
✅ | ✅ | ✅ |
Sub |
✅ | ✅ | ✅ |
Sum |
✅ | ✅ | ✅ |
Tanh |
✅ | ✅ | ✅ |
Transpose |
✅ | ✅ | ✅ |
TFLite¶
| Operator | aries-rb |
regulus-ra |
regulus-rb |
|---|---|---|---|
ADD |
✅ | ✅ | ✅ |
AVERAGE_POOL_2D |
✅ | ✅ | ✅ |
BATCH_MATMUL |
✅ | ✅ | ✅ |
CONCATENATION |
✅ | ✅ | ✅ |
CONV_2D |
✅ | ✅ | ✅ |
COS |
✅ | ❌ | ✅ |
DEPTHWISE_CONV_2D |
✅ | ✅ | ✅ |
DIV |
✅ | ✅ | ✅ |
ELU |
✅ | ✅ | ✅ |
EXP |
✅ | ✅ | ✅ |
EXPAND_DIMS |
✅ | ✅ | ✅ |
FULLY_CONNECTED |
✅ | ✅ | ✅ |
GELU |
✅ | ✅ | ✅ |
HARD_SWISH |
✅ | ✅ | ✅ |
LEAKY_RELU |
✅ | ✅ | ✅ |
LOG |
✅ | ✅ | ✅ |
LOGISTIC |
✅ | ✅ | ✅ |
MAX_POOL_2D |
✅ | ✅ | ✅ |
MAXIMUM |
✅ | ✅ | ✅ |
MEAN |
✅ | ✅ | ✅ |
MINIMUM |
✅ | ✅ | ✅ |
MUL |
✅ | ✅ | ✅ |
POW |
✅ | ✅ | ✅ |
PRELU |
✅ | ✅ | ✅ |
REDUCE_MAX |
✅ | ✅ | ✅ |
REDUCE_MIN |
✅ | ✅ | ✅ |
RELU |
✅ | ✅ | ✅ |
RELU6 |
✅ | ✅ | ✅ |
RELU_0_TO_1 |
✅ | ✅ | ✅ |
RELU_N1_TO_1 |
✅ | ✅ | ✅ |
RESHAPE |
✅ | ✅ | ✅ |
RESIZE_BILINEAR |
✅ | ✅ | ✅ |
RESIZE_NEAREST_NEIGHBOR |
✅ | ✅ | ✅ |
SIN |
✅ | ❌ | ✅ |
SLICE |
✅ | ✅ | ✅ |
SOFTMAX |
✅ | ❌ | ✅ |
SPLIT |
✅ | ✅ | ✅ |
SPLIT_V |
✅ | ✅ | ✅ |
SQRT |
✅ | ✅ | ✅ |
SQUARE |
✅ | ✅ | ✅ |
SQUARED_DIFFERENCE |
✅ | ✅ | ✅ |
SQUEEZE |
✅ | ✅ | ✅ |
STRIDED_SLICE |
✅ | ✅ | ✅ |
SUB |
✅ | ✅ | ✅ |
SUM |
✅ | ✅ | ✅ |
TANH |
✅ | ✅ | ✅ |
TRANSPOSE |
✅ | ✅ | ✅ |
TRANSPOSE_CONV |
✅ | ✅ | ✅ |
Mobilint IR¶
For Mobilint IR operation inputs, attributes, outputs, and formulas, see Mobilint IR Operation Formats.