MagicNET
Cloud Computing & Edge Computing
MagicNet is a neural network computing technology based on heterogeneous computing architecture,
providing high-performance & power-efficient neural network computing for both cloud & edge.
MagicNet has significant performance advantage against mainstream industry standard in both inferencing & training.
MagicNet is able to detect massive small-sized objects in ultra-high network resolution,
which enables brand new possibilities for public security related industries.
providing high-performance & power-efficient neural network computing for both cloud & edge.
MagicNet has significant performance advantage against mainstream industry standard in both inferencing & training.
MagicNet is able to detect massive small-sized objects in ultra-high network resolution,
which enables brand new possibilities for public security related industries.
Significant Neural Network Computing Performance
199X Faster on CPU, 3.5X Faster than GPU
MagicNet is 199 times faster than Intel MKL, and comparing with NVIDIA Titan X workstations,
MagicNet is 3.5X faster with 1/3 power consumption.
MagicNet is 3.5X faster with 1/3 power consumption.
* Comparing items working on the same dataset with the same mAP.
|
|
Detecting Over 2,000 Faces from 4K Camera
Traditional deep learning technology is very weak at detecting tiny objects in high density at the same time,
MagicNet is able to detect over 2,000 objects in a single frame.
MagicNet Face 4K works on 4096 x 2048 network resolution (the neurons see 8.3 mega pixels in a single frame)
while main stream object detectors still work on 300x300 - 512x512 low resolutions.
MagicNet Face 4K can detect small faces (less than 10 x 10 pixels) in a picture that contains thousands of people.
MagicNet is able to detect over 2,000 objects in a single frame.
MagicNet Face 4K works on 4096 x 2048 network resolution (the neurons see 8.3 mega pixels in a single frame)
while main stream object detectors still work on 300x300 - 512x512 low resolutions.
MagicNet Face 4K can detect small faces (less than 10 x 10 pixels) in a picture that contains thousands of people.
High Performance / Watt, Low Deploy & Operating Cost
Ultra-high performance per watt, boosting computing efficiency
MagicNet has very high extreme performance, but for edge computing, MagicNet still works like magic.
MagicNet shows high-end GPU level performance on a 0.9GHz mobile level Intel M3-6Y30 system which cost only 5.9 watt.
This brings new imagination for edge computing – running sophisticated AI model on devices like drones, robots, and cars.
In the cloud, MagicNet works without expensive GPU chips, significantly lowered the cost of AI server deployment.
And with only 1/3 power consumption, MagicNet is a money-saver for the operating cost.
MagicNet shows high-end GPU level performance on a 0.9GHz mobile level Intel M3-6Y30 system which cost only 5.9 watt.
This brings new imagination for edge computing – running sophisticated AI model on devices like drones, robots, and cars.
In the cloud, MagicNet works without expensive GPU chips, significantly lowered the cost of AI server deployment.
And with only 1/3 power consumption, MagicNet is a money-saver for the operating cost.
Performance on VOC 2007+2012 Dataset with same mAP |
AI Model |
Tiny Yolo NVIDIA Optimized CPU + GPU ( i7-7700K + Titan X ) |
Tiny Magic Yolo PQ Labs Optimized CPU Only (0.9Ghz Mobile Chip) |
Frame Per Second |
207 |
157 |
Typical Power Consumption |
400 Watt |
5.9 Watt |
Performance / Watt |
0.52 |
26.61 51X better power efficiency |
