EN
USB Neural Network Computing Card
The deep neural network accelerator based on the artificial intelligence processor SPR2801S is used in the field of high-performance edge computing and can be used as visual-based deep learning operation and AI algorithm acceleration. Universal USB interface for more convenient access to a variety of devices.
Features
- Support USB2.0 and 3.0 standard interface communication
- No programming needed, no language barriers
- Open SDK, can be applied to platforms such as X86, ARM, etc.
- Support Android, Linux and other operating systems
- Support VGG, SSD and other neural network models
Specification
| NPU | |
| Name | Lightspeeur SPR2801S(28nm process, unique MPE and ApiM architecture) |
| Energy efficiency | 9.3 TOPs/Watt |
| Peak | 5.6 Tops@100MHz |
| Low Power | 2.8 Tops@300mW |
| Hardware interface | SDIO3.0 \ eMMC 4.5 |
| Package | BGA(7mm*7mm) |
| Manufacturing process | 28nm |
| USB accelerator | |
| Size | 66x19.5x10mm |
| Interface | USB 2.0,USB3.0 Type-A |
| Transmission Bandwidth | read bandwidth = 68.00 MB/s, write bandwidth = 84.69 MB/s |
| Working Voltag | DC 5V 200mA |
| Operation Temperature | 0° C to 40° C |
| Storage Temperature | -20° C to 80° C |
| Framework | support Pytorch, Caffe framework, follow-up support TensorFlow |
| SDK Provided | ARM、X86 SDK |
| Tools | PLAI model traning tool(support for GG1,GNet18 and GNetfc network models based on VGG-16) Support Ubuntu, Windows operating system |
About Lightspeeur® 2801S
Lightspeeur® 2801S is the world’s first commercially available deep learning CNN accelerator chip to run audio and video processing to power AI devices from Edge to Data Center.
Lightspeeur® pairs with a host processor to improve AI performance, while significantly reducing energy costs by minimizing host processing and power requirements with no extra memory requirements.
Lightspeeur® 2801S uses 100% proprietary and patented technologies to accelerate CNN processing at extremely high speeds, while consuming very little power.
GTI’s Matrix Processing Engine (MPE™) architecture is a multi-dimensional processing array of physical matrices of digital multiply-add (MAC) units that computes the series of matrix operations of a convolutional neural network. The scalable matrix design of the engines allows each engine to directly communicate and interact with adjacent engines, optimizing and accelerating data flow.
Application
- Edge computing
- Intelligent monitoring
- Smart toys and robots
- Smart home
- Virtual reality and augmented reality
- Face detection and recognition
- Speech Recognition
- Natural language processing
- Embedded deep learning device
- Cloud Machine Learning and Deep Learning System
- Artificial intelligence data center server
- Advanced assisted driving and autonomous driving