Akida Pico capabilities

Note

These details are relevant to Akida Pico 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 Pico IP.

Please refer to Akida Pico layers for layer descriptions.

StatefulRecurrent

Model structure

 A model cannot have more than 8
StatefulRecurrent layers.
 Only a Dequantizer or a PicoPostProcessing
layer is allowed at the end of a series of StatefulRecurrent layers, and it must be
the last layer of the model.
 When the ReLU activation is present, it must be
unbounded.

Input

Layer	Input bitwidth	Channels
First	8	<=256
First	16	<=128
Intermediate	8	<=256

For the first layer, the packed input row (channels × input bits) must occupy a power-of-2 number of 32-bit words and cannot exceed 512 bytes.

For example:

Input configuration	Packed size	32-bit words	Status
3 × 8-bit	24 bits	1	Valid
9 × 8-bit	72 bits	3	Mapping error, pad channels to 13-16 (4 words)
3 × 16-bit	48 bits	2	Valid
6 × 16-bit	96 bits	3	Mapping error, pad channels to 7-8 (4 words)

Padding is not performed automatically by MetaTF tools: if the packed input row does not occupy a power-of-2 number of 32-bit words, mapping fails and the user must pad the input channels to reach the next power-of-2 word count.

Parameters

Stateful channels	Output channels	Subsample
<=256, power of 2	<=256	<=4

Stateful channels must be the same for all StatefulRecurrent layers of the model.

The last StatefulRecurrent layer of the model must have 8-bit, 28-bit or 32-bit outputs.

PicoPostProcessing

Input dimensions

Channels	Timesteps
1	[3:128]

Note

In addition to the limits above, the memory required by each layer (events, states, filters and time constants) is checked against the SRAM sizes of the target device at mapping time.