Two methods of quantization supported, Post-training quantization for rapid floating point to fix point trial and evaluation (PTQ), and Quantization-aware training for near zero accuracy loss optimization (QAT).

SDK with abundant documentation, reference samples applications and  reference algorithm implementation, for automotive use cases to accelerate your software deployment.

Model compiler automatically performs optimization such as kernel or feature map level splitting, vertical and horizontal layer fusion and  instruction grouping and scheduling. It increases data or instruction parallelism. It also reduces external memory visits which reduces latency and power consumption.