Model Zoo

Smarter than GPU
Smarter than GPU

As a leading On-Device AI semiconductor company, DEEPX we are proud to introduce our groundbreaking AI quantization
technology, revolutionizing the landscape of computational efficiency and performance.

IQ8, short for Intelligent Quantization Int8, marks significant advancement. Unlike GPU-based solutions using 32 Floating
Point, IQ8 maintains the same level of GPU accuracy or even outperforms GPU accuracy. In comparison with IQ8-Pro,
IQ8-Master can provide the highest model accuracy, but it requires a full dataset with label information for the training.

With our innovative approach, we enable unparalleled optimization of neural networks, delivering significant reductions in
memory usage and computational complexity while preserving model accuracy. Embrace the future of AI processing with our
cutting-edge solutions, driving transformative advancements in technology and paving the way for AI everywhere.

As a leading On-Device AI semiconductor company, DEEPX we are proud to introduce our groundbreaking
AI quantization technology, revolutionizing the landscape of computational efficiency and performance.
IQ8, short for Intelligent Quantization Int8, marks significant advancement.
Unlike GPU-based solutions using 32 Floating Point, IQ8 maintains the same level of GPU accuracy or
even outperforms GPU accuracy. In comparison with IQ8-Pro, IQ8-Master can provide the highest model accuracy,
but it requires a full dataset with label information for the training.
With our innovative approach, we enable unparalleled optimization of neural networks,
delivering significant reductions in memory usage and computational complexity while preserving model accuracy.
Embrace the future of AI processing with our cutting-edge solutions,
driving transformative advancements in technology and paving the way for AI everywhere.

1. Task: Image Classification
Task Model Name Dataset Input Resolution (RGB) # of MACs [G] # of Parameters [M] Accuracy Metric Accuracy DX-M1 FPS
(Batch Size=1)		 Reference
Full Precision
Accuracy (FP32)		 IQ8-Pro
Accuracy (INT8)		 IQ8-Master
Accuracy (INT8)
IC DenseNet121 ImageNet 224X224 3.19 8.04 top1 74.43 72.68 TBU 164.79

IC DenseNet161 ImageNet 224X224 8.43 28.86 top1 77.11 76.41 TBU 88.33

IC EfficientNetB2 ImageNet 288X288 1.60 9.08 top1 80.61 79.19 TBU 286.59

IC EfficientNetV2S ImageNet 384X384 9.47 21.38 top1 84.24 81.30 TBU 155.27

IC HarDNet39DS ImageNet 224X224 0.44 3.48 top1 72.08 71.41 TBU 611.43

IC HarDNet68 ImageNet 224X224 4.26 17.56 top1 76.47 76.34 TBU 187.73

IC MobileNetV1 ImageNet 224X224 0.58 4.22 top1 69.49 68.84 TBU 1140.47

IC MobileNetV2 ImageNet 224X224 0.32 3.49 top1 72.15 71.97 72.76 900.20

IC RegNetX400MF ImageNet 224X224 0.42 5.48 top1 74.88 74.46 TBU 472.14

IC RegNetX800MF ImageNet 224X224 0.81 7.24 top1 77.52 77.26 TBU 356.47

IC RegNetY200MF ImageNet 224X224 0.21 3.15 top1 70.36 70.13 TBU 786.93

IC RegNetY400MF ImageNet 224X224 0.41 4.33 top1 75.78 75.38 TBU 529.49

IC RegNetY800MF ImageNet 224X224 0.85 6.42 top1 78.83 78.54 TBU 334.14

IC ResNeXt26_32x4d ImageNet 224X224 2.49 15.37 top1 75.85 75.68 TBU 265.75

IC ResNeXt50_32x4d ImageNet 224X224 4.27 25.00 top1 81.19 80.95 TBU 159.39

IC ResNet18 ImageNet 224X224 1.82 11.69 top1 69.75 69.60 70.85 833.84

IC ResNet34 ImageNet 224X224 3.67 21.79 top1 73.29 73.27 73.71 441.70

IC ResNet50(v1) ImageNet 224X224 4.12 25.53 top1 75.90 75.72 76.71 370.14

IC ResNet50(v2) ImageNet 224X224 4.12 25.53 top1 80.85 80.66 80.92 370.14

IC ResNet101 ImageNet 224X224 7.84 44.50 top1 81.90 81.62 TBU 221.16

IC SqueezeNet1_0 ImageNet 224X224 0.83 1.25 top1 58.09 57.07 TBU 443.27

IC SqueezeNet1_1 ImageNet 224X224 0.36 1.24 top1 58.18 57.60 TBU 1321.00

IC VGG11BN ImageNet 224X224 7.63 132.86 top1 70.37 70.24 71.62 113.02

IC VGG19BN ImageNet 224X224 19.69 143.67 top1 74.24 74.09 TBU 81.33

IC WideResNet101_2 ImageNet 224X224 22.81 126.82 top1 82.52 82.30 TBU 85.00

IC WideResNet50_2 ImageNet 224X224 11.43 68.85 top1 81.61 81.54 TBU 153.86

IC AlexNet ImageNet 224X224 0.72 61.10 top1 56.56 56.53 57.51 314.56

IC VGG11 ImageNet 224X224 7.63 132.86 top1 69.03 68.94 69.84 118.12

IC VGG13 ImageNet 224X224 11.34 133.05 top1 69.93 69.83 70.86 106.04

IC VGG13BN ImageNet 224X224 11.34 133.05 top1 71.55 71.55 72.70 106.04

IC MobileNetV3Small ImageNet 224X224 0.06 2.54 top1 67.66 TBU 67.50 1644.74

IC MobileNetV3Large ImageNet 224X224 0.23 5.47 top1 75.27 72.82 75.27 844.59

IC OSNet0_5 ImageNet 224X224 0.44 1.14 top1 68.26 67.37 69.01 402.09

IC OSNet0_25 ImageNet 224X224 0.14 0.71 top1 58.88 54.19 59.28 429.74

IC RepVGGA1 ImageNet 320X320 4.83 12.79 top1 75.28 73.69 76.00 450.86

2. Task: Object Detection
Task Model Name Dataset Input Resolution (RGB) # of MACs [G] # of Parameters [M] Accuracy Metric Accuracy DX-M1 FPS
(Batch Size=1)		 Reference
Full Precision
Accuracy (FP32)		 IQ8-Pro
Accuracy (INT8)		 IQ8-Master
Accuracy (INT8)
OD SSDMV1 PascalVOC 300X300 1.55 9.48 mAP50 67.59 67.58 TBU 531.87
OD SSDMV2Lite PascalVOC 300X300 0.70 3.38 mAP50 68.70 68.73 69.52 450.33
OD YoloV3 COCO 640X640 81.13 62.02 mAP50:95 46.65 46.41 TBU 31.87
OD YoloV5N COCO 640X640 2.71 1.97 mAP50:95 28.08 27.01 28.26 183.48
OD YoloV5S COCO 640X640 9.10 7.33 mAP50:95 37.45 36.91 37.36 156.91
OD YoloV5M COCO 640X640 26.07 21.27 mAP50:95 45.08 44.67 45.07 66.99
OD YoloV5L COCO 640X640 57.10 46.64 mAP50:95 48.74 48.34 47.72 41.85
OD YoloV7 COCO 640X640 55.28 36.92 mAP50:95 50.86 50.69 TBU 39.12
OD YoloV7E6 COCO 1280X1280 269.21 97.27 mAP50:95 55.22 55.15 TBU 7.83
OD YoloV7Tiny COCO 640X640 7.01 6.24 mAP50:95 37.29 37.08 TBU 137.18
OD YOLOX_S COCO 640X640 14.41 8.96 mAP50:95 40.45 40.17 40.30 106.15
OD YOLOv8L COCO 640X640 85.13 43.69 mAP50:95 52.75 52.14 52.77 26.24

3. Task: Segmentation
Task Model Name Dataset Input Resolution (RGB) # of MACs [G] # of Parameters [M] Accuracy Metric Accuracy DX-M1 FPS
(Batch Size=1)		 Reference
Full Precision
Accuracy (FP32)		 IQ8-Pro
Accuracy (INT8)		 IQ8-Mater
Accuracy (INT8)
SEG BiSeNetV1 CITY 1024X2048 118.98 13.27 mIOU 75.37 74.67 TBU 19.03
SEG BiSeNetV2 CITY 1024X2048 99.14 3.35 mIOU 74.95 74.54 75.00 17.74
SEG DeepLabV3PlusMobilenet VOC2012 512X512 26.62 5.80 mIOU 68.48 68.22 TBU 84.63

4. Task: Face ID
Task Model Name Dataset Input Resolution (RGB) # of MACs [G] # of Parameters [M] Accuracy Metric Accuracy DX-M1 FPS
(Batch Size=1)		 Reference
Full Precision
Accuracy (FP32)		 IQ8-Pro
Accuracy (INT8)		 IQ8-Master
Accuracy (INT8)
FD YOLOv5s_Face WIDERFace (easy) 640X640 8.53 8.073 AP50 94.57 95.08 TBU 139.25
FD YOLOv5s_Face WIDERFace (medium) 640X640 8.53 8.07 AP50 92.94 93.58 TBU 139.25

FD YOLOv5s_Face WIDERFace (hard) 640X640 8.53 8.07 AP50 83.70 84.71 TBU 139.25

FD YOLOv5m_Face WIDERFace (easy) 640X640 25.84 22.00 AP50 95.51 95.31 TBU 62.87

FD YOLOv5m_Face WIDERFace (medium) 640X640 25.84 22.00 AP50 94.03 93.76 TBU 62.87

FD YOLOv5m_Face WIDERFace (hard) 640X640 25.84 22.00 AP50 85.65 85.11 TBU 62.87

FD YOLOv7s_Face WIDERFace (easy) 640X640 9.35 6.26 AP50 94.86 94.67 TBU 108.53

FD YOLOv7s_Face WIDERFace (medium) 640X640 9.35 6.26 AP50 93.30 93.06 TBU 108.53

FD YOLOv7s_Face WIDERFace (hard) 640X640 9.35 6.26 AP50 85.30 85.06 TBU 108.53

FD YOLOv7_Face WIDERFace (easy) 640X640 54.63 38.55 AP50 96.93 96.81 TBU 41.36

FD YOLOv7_Face WIDERFace (medium) 640X640 54.63 38.55 AP50 95.69 95.59 TBU 41.36

FD YOLOv7_Face WIDERFace (hard) 640X640 54.63 38.55 AP50 88.34 88.38 TBU 41.36

FD YOLOv7_TTA_Face WIDERFace (easy) 640X640 54.63 38.55 AP50 96.92 96.86 TBU 41.25

FD YOLOv7_TTA_Face WIDERFace (medium) 640X640 54.63 38.55 AP50 95.69 95.65 TBU 41.25

FD YOLOv7_TTA_Face WIDERFace (hard) 640X640 54.63 38.55 AP50 88.34 88.33 TBU 41.25

FD YOLOv7_W6_Face WIDERFace (easy) 960X960 100.21 74.44 AP50 96.37 96.77 TBU 22.47

FD YOLOv7_W6_Face WIDERFace (medium) 960X960 100.21 74.44 AP50 95.07 95.56 TBU 22.47

FD YOLOv7_W6_Face WIDERFace (hard) 960X960 100.21 74.44 AP50 88.59 88.29 TBU 22.47

FD YOLOv7_W6_TTA_Face WIDERFace (easy) 1280X1280 178.16 77.96 AP50 95.98 96.06 TBU 12.51

FD YOLOv7_W6_TTA_Face WIDERFace (medium) 1280X1280 178.16 77.96 AP50 94.98 95.05 TBU 12.51

FD YOLOv7_W6_TTA_Face WIDERFace (hard) 1280X1280 178.16 77.96 AP50 89.44 89.77 TBU 12.51

5. Task: Image De-noising
Task Model Name Dataset Input Resolution (RGB) # of MACs [G] # of Parameters [M] Accuracy Metric Accuracy DX-M1 FPS
(Batch Size=1)		 Reference
Full Precision
Accuracy (FP32)		 IQ8-Pro
Accuracy (INT8)		 IQ8-Master
Accuracy (INT8)
DN DnCNN_15 BDS68 512X512 145.80 0.56 PSNR 31.72 31.47 TBU 21.65
DN DnCNN_25 BDS68 512X512 145.80 0.56 PSNR 29.20 28.76 TBU 21.78

DN DnCNN_20 BDS68 512X512 145.80 0.56 PSNR 26.21 24.95 TBU 21.70