s73108bit inference with tensorrt.pdf ● Goal: Convert FP32 CNNs into INT8 without significant accuracy loss. ● Why: INT8 math has higher throughput, and lower memory requirements. ● Challe