Akida 1.0 Hardware constraints

Warning

The following constraints concern only Akida 1.0 IP based solutions and the AKD1000 reference SoC.

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 AKD1000 SoC.

Please refer to Akida V1 layers for layers description.

InputConvolutional

Input dimensions

Width	Height	Channels
[5:256]	>= 5	1, 3

Convolution parameters

Kernel Size	Stride	Type
3×3, 5×5, 7×7	1, 2, 3	Same, Valid

Maximum number of filters

Kernel Sise	3×3	5×5	7×7
Max(1 ch)	512	192	96
Max(3 ch)	192	64	32

Max Pooling size

1×1, 1×2, 2×1, 2×2

Pool stride is equal to pool size for InputConvolutional.

Quantization bitwidth

Input	Weights	Activation
8	1, 2, 4, 8	1, 2, 4

While minimum weights bitwidth supported is 1 for native learning, CNN2SNN quantization only allows quantization with bitwidth >=2 because float weights are signed while 1-bit integers are unsigned by definition.

Convolutional

Convolution parameters

Kernel Size	Stride	Type
1×1, 3×3, 5×5, 7×7	1, 2	Same

Stride 2 is only supported with 3×3 kernels.

Max Pooling parameters

Size	Stride
1×1, 2×2	1, 2

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

Global Average Pooling width

[1:32]

The output of the convolution must have at least 3 rows.

Quantization bitwidth

Input	Weights	Activation
1, 2, 4	1, 2, 4	1, 2, 4

While minimum weights bitwidth supported is 1 for native learning, CNN2SNN quantization only allows quantization with bitwidth >=2 because float weights are signed while 1-bit integers are unsigned by definition.

SeparableConvolutional

Convolution parameters

Kernel Size	Stride	Type
3×3, 5×5, 7×7	1, 2	Same

Stride 2 is only supported with 3×3 kernels.

Max Pooling parameters

Size	Stride
1×1, 2×2	1, 2

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

Global Average Pooling width

[1:32]

The output of the convolution must have at least 3 rows, 1×1 inputs are not supported.

Quantization bitwidth

Input	Weights	Activation
1, 2, 4	2, 4	1, 2, 4

While minimum weights bitwidth supported is 1 for native learning, CNN2SNN quantization only allows quantization with bitwidth >=2 because float weights are signed while 1-bit integers are unsigned by definition.

FullyConnected

Input dimensions

Width	Height	WxHxC
1	1	<= 57334

Quantization bitwidth

Input	Weights	Activation
1, 2, 4	1, 2, 4	1, 2, 4

While minimum weights bitwidth supported is 1 for native learning, CNN2SNN quantization only allows quantization with bitwidth >=2 because float weights are signed while 1-bit integers are unsigned by definition.