ApproxiNet
Years :
Subjects :
Laboratories :
The project ApproxiNet was a 2021 winner of the PREMATURATION call for projects, led by the Innovation Lab of IP Paris.
THE PROBLEM ADDRESSED
With Neural Network applications in AI taking a prevalent role in our society, it is strategic to develop specialized hardware. ApproxiNet proposes a Micro-NPU (Neuro-Processing-Unit) targeted at mobile visual and audio applications with a focus on ultra low power consumption.
TECHNOLOGY
- An Application-specific integrated circuit (ASIC) leveraging sparsity and quantization in neural networks through adaptive training
- DCMI interface for video, SPI for audio and IMU. Customized for Edge AI Video/Audio/IMU applications.
- Resnet50 running at 30fps, with ~10ms latency
COMPETITIVE ADVANTAGES
- Easy Integration into Existing ARM based SoCs
- Ultra low power operation (Packing 2 Tops/S in 0.5 mm2 (28nm) with ultra low power operation @10Tops/Watt)
- 8 bits features, and customizable 1-4 bits weights without loss of precision
- Look-Ahead fusion of network layers, resulting in low memory bandwidth, No DDR memory required.
- Seamless integration with Tensorflow-Lite