DEEPX 新款 M1 NPU 提供灵活性、准确性、效率和性能.pdf

1、DEEPXs New M1 NPU Delivers Flexibility,Accuracy,Efficiencyand Performance Jay KimExecutive Vice PresidentDEEPXDisruptive Innovation “ITS REAL”2 2023 DEEPXDX-M1:133 FPS(Yolov7)mAP:50.900V100:229 FPS(Yolov7)mAP:51.0Batch Size=1Batch Size=1DEEPXs Key Differentiators 2023 DEEPX3DEEPXs Key Differentiator

2、s 1.Flexibility 4 2023 DEEPX Latest:Latest DNN algorithms(SOTA)-EfficientNet,YoloX,YoloV7,PIDNet Wide Range:Various kinds of algorithms-Classification,object detection,segmentation,pose estimation,anomaly detection Complex Algorithms:Normally too complex for edge NPUs -DeepX NPU can run PIDNet,SFA3D

3、(Sensor Fusion)Customers can use our NPU longer and wider.Evidence#1:SOTA Support(Object Detection)5 2023 DEEPXDEEPXs Key Differentiators 2.Accuracy Use int 8-bit instead of 32-bit floating point operations-Expect accuracy drop comparing to GPU Almost the same accuracy as a GPU can get-Normally same

4、 or less than 1%accuracy drop expected Wow factor:Better accuracy than GPU in some cases-YoloXs,MobileNetv1,ResNet50,EfficientNet-B0 and etc Customers can get the most accurate result with our NPU.2023 DEEPX6Evidence#2:AI Accuracy ComparisonDEEPX accuracy is better than FP32&other latest NPUs 2023 D

5、EEPX7DEEPXs Key Differentiators 3.Efficiency Popular power efficiency factor(TOPS/W)with Resnet50-World top power efficiency:over 10 TOPS/WPrefer other efficiency factor(FPS/TOPS)-Actual result with 1 TOPS Finally pursue the effective power efficiency such as FPS/W-Actual result with 1 Watt Customer

6、s can get the max performance with the lowest power.2023 DEEPX8Extreme Case:Ultra Low Power NPUCMOS IMAGE SENSORIntelligent CMOS Image SensorImage Enhancement NPU(PoC)Face Recognition Function NPU(PoC)1 mm X 1 mm 40 nm Lower than 10 mW&10 fps Face recognition