nearbAI on ASIC
nearbAI – AI close to the sensors
The nearbAI semiconductor IP is a small, low-power and affordable AI core that runs locally, close to your sensors. This results in a low and predictable latency and runs with ultra low-power consumption. easics’ embedded AI solutions integrate tightly with novel and existing sensors such as image sensors capturing light inside and outside the visible spectrum (such as hyperspectral and thermal infrared), 3D scanning laser (LiDAR), Time-of-Flight (ToF) sensors, radar, microscopy, ultrasound sensors, and microphones.
Why choose nearbAI as AI accelerator?
nearbAI software tools
Supported operations
The core is optimised by parameterisation of our generic core towards application specific needs. nearbAI supports the following operations:
- 2D convolution
- 2D ConvLSTM layers
- depthwise convolutions
- dilated convolutions
- transposed convolutions
- matrix multiplications for LSTM optimization
- fully connected layers
- bias
- max pooling, average pooling
- ReLU, ReLU6, Leaky ReLU
- non-supported operations are developed on demand
- DNN examples:
- CNNs:ResNet, YOLO, MobileNet,..
- RNNs: DeepSpeech, …
Which sensors benefit from nearbAI?
Which customers benefit from nearbAI?
The nearbAI IP offers benefits to semiconductor companies and original equipment manufacturers (OEM).
Semiconductor companies
OEMs
nearbAI architecture
The DMA Controller loads input data and quantized weights into internal buffers. Both data and weights are then shifted through the Convolution Engine. The results are sent to the Accumulator and are finalized in the Post Processor. The Sequencer manages the execution of the subsequent layers of the DNN. It generates a continuous flow of tensors through all layers of the DNN. The final output tensors are returned as results —structured data— to your application.
Deep learning on ASIC – download PDF documentation
Want to know more about nearbAI?
