AndesAIRE™ AnDLA™ I350

A high performance-efficient edge deep learning accelerator for CNN, RNN and LSTM

AndesAIRE™ AnDLA™ I350 Overview

AndesAIRE™ AnDLA™ I350 is a deep learning accelerator (DLA) designed for high-performance, cost-sensitive edge and
endpoint AI inference. It supports popular AI framework formats (TensorFlow Lite, PyTorch, ONNX), NN models (CNN/RNN),
and performs various INT8 operations, including convolution, fully connected, elementwise, activation, pooling, channel
padding, upsampling, and concatenation.

With internal DMA and local memory, AndesAIRE™ AnDLA™ I350 efficiently maximizes hardware utilization. It uses
operator and layer fusion to accelerate common operator sequences, boosting performance and reducing memory access.

Key configurable parameters of AndesAIRE™ AnDLA™ I350 include MAC count and local memory size, delivering flexible
computing power up to 4 TOPS at 1 GHz for diverse applications.

Key Features

  • Configurable MACs: 32, 64, 128, 256, 512, 1024, 2048 (INT8)
  • Maximum performance up to 4 TOPS at 1GHz
  • Configurable local memory: 16KB to 4MB
  • Multi-dimensional Direct Memory Access (DMA)
  • Bus interfaces: AHB-64b, AXI-64/128/256/512b
  • Neural network models:
    • Image and Video: MobileNet-v1/v2, ResNet-8/50, Tiny YOLO v1/v2, YOLO v3/v5, Inception-v2, EfficientNet-Lite, MobileNet-v1-SSD, etc
    • Speech/Voice and Audio: RNNoise, etc
  • Operators: Conv2D, Depthwise Convolution, Pointwise Convolution, Elementwise (Add, Add_const, Sub, Sub_const, Mult, Mult_const), Fully Connected, Activation (Hardsigmoid, Hardswish, leaky ReLU, ReLU, ReLU6, ReLU_n1_to_1, MISH, Softplus, Logistic, Swish, Tanh), Pooling (Average, Global Avg, Max), Upsampling, Concatenation, Split, Strided Slice, Reshape, Transpose, Batch Normalization, Channel Padding, GRU, etc
  • Operator and layer fusion

Applications

  • AIoT and TinyML on edges and endpoints
  • Smart cameras
  • Smart sensors and sensor hubs
  • Wearables and TWS earbuds
  • Smart home appliances
  • Smart cockpit

Block Diagram

Development Tools

  • AndesAIRE™ NN SDK
    • NNPilot™ neural network optimization tool suite
    • LiteRT and LiteRT for Microcontrollers for AnDLA™
    • AnDLA™ driver and runtime
  • AndesAIRE™ NN Library
  • AndeSight™ IDE