Akida 2.0 capabilities

Note

These details are relevant to Akida 2.0 IP-based solutions.

For compatibility with multiple possible hardware backends downstream, CNN2SNN and the Akida simulator impose few constraints on layer dimensions. However, hardware does have limits in this respect, which will be checked at the stage of mapping a model to a specific device (real or virtual). This page details the limits for the Akida 2.0 IP.

Please refer to Akida V2 layers for layers description.

Note

Akida 2.0 supports using a Lookup Table (LUT) for activation functions other than ReLU for InputConv2D, Conv2D, DepthwiseConv2D, Conv2DTranspose and DepthwiseConv2DTranspose layers. However, Max Pooling is not compatible when LUT-based activations are used.

Input dimensions

Width

Height

Channels

>=5

[5:384]

1, 3

Input shape is (Width, Height, Channels).

Convolution parameters

Kernel Size

Stride

Type

3×3

1, 2, 3

Same, Valid

4×4

1, 2, 3, 4

Same, Valid

5×5

1, 2, 3, 4, 5

Same, Valid

7×7

1, 2, 3, 4, 5, 6

Same, Valid

512 filters max.

Max Pooling size

(size, stride) supported combinations: (2, 2), (3, 2).

Quantization bitwidth

Input

Weights

Activation

8

8

4, 8

Input dimensions

Width

Height

Channels

<=4096

<=4096

<= 2048

Convolution parameters

Kernel Size

Stride

Type

1×1, 3×3, 5×5, 7×7

1, 2

Same
1×1 kernel is only available for Conv2D.
Stride 2 is only supported with 3×3 kernels.

Max Pooling parameters

Size

Stride

2×2

1, 2

Pooling stride cannot be greater than pooling size, layer with max pooling must be followed by another Conv2D or DepthwiseConv2D layer.

Global Average Pooling width

[1:64], Width x Height < 144

The output of the convolution must have at least 1 rows.

Quantization bitwidth

Input

Weights

Activation

4, 8

4, 8

4, 8

Input dimensions

Width

Height

Channels

<=4096

<=4096

<= 2048

Convolution parameters

Kernel Size

Stride

Type

3×3, 4×4

2

Same

Quantization bitwidth

Input

Weights

Activation

8

8

8

Input dimensions

Width

Height

WxHxC

1

1

<= 2048

Quantization bitwidth

Input

Weights

Activation

4, 8

4, 8

4, 8

Input dimensions

Width

Height

Channels

<=4096

<=4096

<= 2048

Parameters

FIFO size

Filters

[2:10]

<=2048

Quantization bitwidth

Input

Weights

Activation

8

8

8