Implementing AI/ML with Lattice sensAI

Introduction

lattice sensAI provides a complete solution stack for implementing AI/ML inference on low-power FPGAs. This guide demonstrates how to convert and deploy a neural network model.

sensAI Components

  • Neural Network Compiler: Converts trained models to FPGA-optimized implementations
  • IP Cores: Pre-optimized acceleration engines
  • Reference Designs: Complete working examples

    • Model Preparation

    Supported Frameworks

    • TensorFlow
    • PyTorch
    • ONNX
    • Caffe

    Model Requirements

    • Quantized to 8-bit weights and activations
    • Fixed input dimensions
    • Supported layer types: Conv2D, Dense, MaxPool, ReLU, etc.

    Implementation Steps

    Step 1: Model Conversion

    Use sensAI Studio to import and optimize your model:

  • Launch sensAI Studio
  • Import your trained model
  • Configure quantization settings
  • Run optimization
  • Export FPGA-compatible model

    • Step 2: Generate FPGA Design

    ``python

    Example: Generate inference engine

    from sensai import ModelCompiler

    compiler = ModelCompiler() compiler.load_model('my_model.onnx') compiler.set_target_device('LFD2NX-40') compiler.compile() compiler.generate_ip_core('my_model_ip') ``

    Step 3: Integrate with Your Design

  • Add the generated IP core to your Radiant project
  • Connect input/output interfaces
  • Add pre/post-processing logic

    • Step 4: Power Optimization

    • Use dynamic frequency scaling
    • Implement clock gating
    • Leverage FD-SOI body biasing

    Performance Benchmarks

    ModelDeviceFPSPower (mW)
    MobileNet v2LFD2NX-4030150
    ResNet-18LFD4NX-10015280
    Tiny YOLOLFD4NX-10025320

    Best Practices

  • Start with reference designs
  • Profile model layer by layer
  • Optimize memory access patterns
  • Use batch processing when possible