latest

Release with SDK version 1.6.6

latest

TVM AOT Migration – Release Notes

Architectural Changes

🧠 ML Inference Backend

• Previous Implementation: TensorFlow Lite Micro (TFLM)
• New Implementation: TVM AOT–compiled inference module
• Benefits:
  • No interpreter or dynamic operator registry
  • Fully static execution graph
  • Lower memory footprint and reduced runtime overhead

Performance

⚡ Hardware Optimization

• Replaced: NEON-optimized TFLM kernels
• Added: TVM-generated fused operators with compile-time scheduling
• Result:
  • Faster inference on 96×96 ROIs
  • Lower inference latency per frame
  • More predictable performance due to static memory layout

📊 Overall Improvements

• ✅ Reduced RAM usage (no tensor arena interpreter)
• ✅ Faster execution from TVM operator fusion + scheduling
• ✅ Smaller binary (unused operators removed at compile time)
• ✅ Improved CPU efficiency on low-power ARM devices

Technical Details

• Model Format: TVM AOT C module
• Execution Model: Direct call to a statically generated entrypoint
• Memory Model: Fully static buffers, no dynamic allocations
• Acceleration: NEON SIMD preprocessing remains (resize, grayscale, Sobel, etc.)