Unusual engine size when generated on JetPack 4.6.4

[hanjen.ai.dev]

Hi,
When generating TRT engine file, the file size is quite weird on one of my model on xavier running jetpack 4.6.4. here are some of the numbers:
original onnx file: 6.2M
and the engine files generated:
RTX 3080 mobile: 5.0M
orin nano dev running jetpack 5.1.2: 5.7M
xavier nx running jetpack 4.5.1: 3.4M
xavier nx running jetpack 4.6.4: 61.2M
xavier nx running jetpack 5.1.2: 5.4M
the above engines are generated with optimization profile set to min =1x1x64x128, opt=8x1x64x128, max=24x1x64x128 and fp16

to further experiment with the settings, set optimization profile to the same size 16x1x64x128 with fp16, the generated engine size changed to
xavier nx running jetpack 4.6.4: 12.4M

I would like to know what is wrong with this model, here is the model that I am having problem with.
Thanks.

Re_20230829_op16-sim.onnx (5.9 MB)

[AastaLLL]

Hi,

Since TensorRT picks different algorithms based on the hardware/software resources.
So it’s possible that it picks some algorithm that contains larger precompute data.

If the inference output runs okay, it should be fine.

Thanks.