Model zoo performance

The Brainchip akida_models package offers a set of pre-built Akida-compatible models (e.g. MobileNet, AkidaNet), pretrained weights for those models and training scripts. Please refer to the model zoo API reference for a complete list of the available models.

This page lists the performance of all models from the zoo reported for Akida 1.0, Akida 2.0 and Akida Pico. Please refer to:

• Akida 1.0 models for models targeting the Akida Neuromorphic Processor IP 1.0 and the AKD1000 reference SoC,

• Akida 2.0 models for models targeting the Akida Neuromorphic Processor IP 2.0,

• Akida Pico models for models targeting the Akida Pico Neuromorphic Processor IP,

• Upgrading to Akida 2.0 tutorial to understand the architectural differences between 1.0 and 2.0 models and their respective workflows.

Note

The download links provided point towards standard TensorFlow Keras models that must be converted to an Akida model using cnn2snn.convert.

Akida 1.0 models

For 1.0 models, 4-bit accuracy is provided and is always obtained through a QAT phase.

Note

• The “8/4/4” quantization scheme stands for 8-bit weights in the input layer, 4-bit weights in other layers and 4-bit activations.

• The NPs column provides the minimal number of neural processors required for the model execution on the Akida IP. The numbers given are the result of the map operation using the Minimal MapMode targeting AKD1000/AKD1500 SoC.

• Energy per inference is an average, measured on an AKD1500 device for the most efficient mapping.

Image domain

Classification

Architecture	Resolution	Dataset	#Params	Quantization	Top-1 accuracy	Size (KB)	NPs	Energy (mJ)	Download
AkidaNet 0.25	160	ImageNet	480K	8/4/4	42.58%	403.3	20	1.56
AkidaNet 0.5	160	ImageNet	1.4M	8/4/4	57.80%	1089.1	24	4.32
AkidaNet	160	ImageNet	4.4M	8/4/4	66.94%	4061.1	68	13.66
AkidaNet 0.25	224	ImageNet	480K	8/4/4	46.71%	409.1	22	2.50
AkidaNet 0.5	224	ImageNet	1.4M	8/4/4	61.30%	1202.2	32	6.885
AkidaNet	224	ImageNet	4.4M	8/4/4	69.65%	6294.0	116	26.34
AkidaNet 0.5 edge	160	ImageNet	4.0M	8/4/4	51.66%	2017.4	38	6.63
AkidaNet 0.5 edge	224	ImageNet	4.0M	8/4/4	54.03%	2130.5	46	9.99
AkidaNet 0.5	224	PlantVillage	1.1M	8/4/4	97.92%	1019.1	33	7.17
AkidaNet 0.25	96	Visual Wake Words	229K	8/4/4	84.77%	179.6	16	0.49
MobileNetV1 0.25	160	ImageNet	467K	8/4/4	36.05%	376.4	20	1.53
MobileNetV1 0.5	160	ImageNet	1.3M	8/4/4	54.59%	1007.0	24	4.21
MobileNetV1	160	ImageNet	4.2M	8/4/4	65.47%	3525.8	65	13.44
MobileNetV1 0.25	224	ImageNet	467K	8/4/4	39.73%	377.9	22	2.46
MobileNetV1 0.5	224	ImageNet	1.3M	8/4/4	58.50%	1065.3	32	6.68
MobileNetV1	224	ImageNet	4.2M	8/4/4	68.76%	5223.3	110	26.28
GXNOR	28	MNIST	1.6M	2/2/1	98.03%	412.8	3	0.34

Object detection

Architecture	Resolution	Dataset	#Params	Quantization	mAP	Size (KB)	NPs	Energy (mJ)	Download
YOLOv2	224	PASCAL-VOC 2007 - person and car classes	3.6M	8/4/4	41.51%	3061.4	71	14.61
YOLOv2	224	WIDER FACE	3.5M	8/4/4	77.63%	3053.1	71	14.22

Regression

Architecture	Resolution	Dataset	#Params	Quantization	MAE	Size (KB)	NPs	Energy (mJ)	Download
VGG-like	32	UTKFace (age estimation)	458K	8/2/2	6.1791	138.6	6	0.14

Face recognition

Architecture	Resolution	Dataset	#Params	Quantization	Accuracy	Size (KB)	NPs	Energy (mJ)	Download
AkidaNet 0.5	112×96	CASIA Webface face identification	2.3M	8/4/4	70.18%	1930.1	21	3.73
AkidaNet 0.5 edge	112×96	CASIA Webface face identification	23.6M	8/4/4	71.13%	6980.2	34	8.41

Audio domain

Keyword spotting

Architecture	Dataset	#Params	Quantization	Top-1 accuracy	Size (KB)	NPs	Energy (mJ)	Download
DS-CNN	Google Speech Commands	22.7K	8/4/4	91.72%	23.1	5	0.07

Point cloud

Classification

Architecture	Dataset	#Params	Quantization	Accuracy	Size (KB)	NPs	Download
PointNet++	ModelNet40 3D Point Cloud	602K	8/4/4	79.78%	490.9	12

Akida 2.0 models

For 2.0 models, both 8-bit PTQ and 4-bit QAT numbers are given. When not explicitly stated, 8-bit PTQ accuracy is given as is (i.e. no further tuning/training, only quantization and calibration). The 4-bit QAT is the same as for 1.0.

Note

• The digit in the quantization scheme stands for both the weights and activations bitwidth. Weights in the first layer are always quantized to 8-bit.

• The NPs column provides the minimal number of neural processors required for the model execution on the Akida IP. The numbers given are the result of the map operation using the Minimal MapMode targeting a 12-node Akida 2.0 device.

Image domain

Classification

Architecture	Resolution	Dataset	#Params	Quantization	Accuracy	NPs	Download
AkidaNet 0.25	160	ImageNet	483K	8 4	48.61% 40.69%	27 26
AkidaNet 0.5	160	ImageNet	1.4M	8 4	61.92% 57.42%	42 29
AkidaNet	160	ImageNet	4.4M	8 4	69.96% 66.80%	124 60
AkidaNet 0.25	224	ImageNet	483K	8 4	52.38% 44.48%	31 26
AkidaNet 0.5	224	ImageNet	1.4M	8 4	64.85% 60.53%	55 34
AkidaNet	224	ImageNet	4.4M	8 4	72.23% 69.21%	206 92
AkidaNet 0.5	224	PlantVillage	1.2M	8 4	99.61% 99.30%	56 35
AkidaNet 0.25	96	Visual Wake Words	227K	8 4	87.05% 85.70%	25 24
AkidaNet18	160	ImageNet	2.4M	8	64.72%	61
AkidaNet18	224	ImageNet	2.4M	8	67.32%	86
MobileNetV1 0.25	160	ImageNet	469K	8 4	45.72% 36.96%	31 30
MobileNetV1 0.5	160	ImageNet	1.3M	8 4	60.16% 54.09%	47 34
MobileNetV1	160	ImageNet	4.2M	8 4	69.04% 64.92%	114 68
MobileNetV1 0.25	224	ImageNet	469K	8 4	49.58% 40.80%	36 31
MobileNetV1 0.5	224	ImageNet	1.3M	8 4	63.67% 57.87%	65 44
MobileNetV1	224	ImageNet	4.2M	8 4	71.31% 67.72%	184 106
GXNOR	28	MNIST	1.6M	4	98.57%	4

Object detection

Architecture	Resolution	Dataset	#Params	Quantization	mAP 50	NPs	Download
YOLOv2 (AkidaNet 0.5 backbone)	224	PASCAL-VOC 2007	3.6M	8 4	51.41% 46.74%	119 70
CenterNet (AkidaNet18 backbone)	384	PASCAL-VOC 2007	2.4M	8	72.77% [1]	336
CenterNet (AkidaNet18 backbone)	224	PASCAL-VOC 2007	2.4M	8	66.08% [1]	125
YOLOv2 (AkidaNet 0.5 backbone)	224	WIDER FACE	3.6M	8 4	80.51% 78.69%	117 69

[1] (1,2)

PTQ accuracy boosted with 1 epoch QAT.

Regression

Architecture	Resolution	Dataset	#Params	Quantization	MAE	NPs	Download
VGG-like	32	UTKFace (age estimation)	458K	8 4	6.0299 6.1421	7 6

Face recognition

Architecture	Resolution	Dataset	#Params	Quantization	Accuracy	NPs	Download
AkidaNet 0.5	112×96	CASIA Webface face identification	2.3M	8 4	73.02% 68.60%	40 29

Segmentation

Architecture	Resolution	Dataset	#Params	Quantization	Binary IOU	NPs	Download
AkidaUNet 0.5	128	Portrait128	1.1M	8	0.9076 [2]	66

[2]

PTQ accuracy boosted with 1 epoch QAT.

Audio domain

Keyword spotting

Architecture	Dataset	#Params	Quantization	Top-1 accuracy	NPs	Download
DS-CNN	Google Speech Commands	23.8K	8 4	92.83% 92.58%	9 9

Point cloud

Classification

Architecture	Dataset	#Params	Quantization	Accuracy	NPs	Download
PointNet++	ModelNet40 3D Point Cloud	277K	8 4	79.62% [3] 79.50%	13 11

TENNs

Gesture recognition

Dataset	#Params	Quantization	Accuracy	NPs	Download
DVS128	165K	8	97.12%	25
Jester	1.3M	8	95.04%	43

Eye tracking

Dataset	#Params	Quantization	Accuracy	NPs	Download
Eye tracking CVPR 2024	219K	8	p10: 98.58% mean_distance: 2.17	22

[3]

PTQ accuracy boosted with 5 epochs QAT.

Akida Pico models

Pico models are recurrent TENNs targeting the Akida Pico Neuromorphic Processor IP. Please refer to the Recurrent TENNs API for model descriptions and to the Akida Pico layers hardware constraints for mapping limits.

Audio domain

Keyword spotting

Architecture	Dataset	#Params	Quantization	Accuracy	Download
TENN recurrent	Google Speech Commands	46.6K	8	93.80%

Vibration domain

Fault classification

Architecture	Dataset	#Params	Quantization	Avg AUROC	Download
TENN recurrent	UORED-VAFCLS	16.6K	8	0.9420