igpu (Tegra) on px2 support for tensorrt int8 (Tensorrt 3.0.2）

Autonomous Vehicles DRIVE PX 2 General
1

Hi,

I have installed the latest version driveworks (5.0.5.0a (February 12, 2018)). The Tensorrt 3.0.2 works fine with dgpu (CUDA_VISIBLE_DEVICES=0), but showing “hardware without native INT8 support” with igpu running (CUDA_VISIBLE_DEVICES=1).My code runs correctly with igpu but the speed is too slow with 48ms/image on igpu comparing 6.6ms/image on dgpu.

I am wondering if the speed issue is due to igpu hasn’t suppport tensorrt3.0.2 int8 ? or is there any other issues out there?

Thank,
Lei

2

Dear lshi,
iGPU supports FP32,FP16 and dGPU supports FP32,INT8.

3

Hi Siva,

Thanks for replying to me. I am wondering if there is a plan to let iGPU to support Int8 in the future tensorRT release? If so, when do you think it will happen?

Thanks,
Lei

4

Hi lshi,
TensorRT does not support it as there is no native hardware support on iGPU.

5

@SivaRamaKrishna Are there any samples for int8 inference using a .uff file on PX2?