The impact of multi-batch int8 quantization on engine size and acc

[HollrayChan]

hi，when I quantized the int8 model on the A10 T4 A2, I found the following situation：

platform  |  batchsize  |   engine size |   acc   |
A10         |         1         |         93M     |    95% |
A10         |         8        |          69M     |   95%  |
T4            |        1         |         69M     |   95%  |
T4            |        8         |         69M     |   95%  |
A2            |        1         |         69M     |   60% |
A2            |        8         |         69M     |   2%   |

My environment：

trt TensorRT-8.2.1.8
cuda 11.5
Operating System: centos 8
pytorch: 1.10.2+cu113
Python:3.6.8
Baremetal or Container (if so, version): Docker

My question：

1.I tried to use torch2trt for conversion, I also tried to use onnx and trtexec, the result is still the same as in the table, why does the engine become larger on the A10？Is there an api or parameter in trt to adjust this situation？

2.The size of the A2 engine seems to be normal, but the model acc declines seriously. I used ENTROPY_CALIBRATION_2 and the same data to calibrate the above test. The MINMAX_CALIBRATION calibration is also the same, and the acc will be worse. I don't know whether the A2 is broken, in the case of fp16, the performance of the graphics card is normal, I don't know what is causing it，or the graphics card of the Ampere architecture needs additional parameters.

Please help me with these two problems, Thanks a lot!