read the article and I would agree with that conclusion; they are saying that if you train, knowing the target is 8bit int instead of 32fp, then the error in downresing isnt' so bad. if you didn't train with that in mind, its not acceptable.From what I read it's the training that has changed to account for int8 (which has been there for I don't know how long), maybe verygreen knows.
The OP also says this
Improving INT8 Accuracy Using Quantization Aware Training and the NVIDIA TAO Toolkit | NVIDIA Developer BlogDeep neural network (DNN) models are routinely used in applications requiring analysis of video stream content. These may include object detection, classification, and segmentation. Typically…developer.nvidia.com