阅读原文请点击
摘要：# 0. 简介 Jetson TX2【1】是基于 NVIDIA Pascal™ 架构的 AI 单模块超级计算机，性能强大（1 TFLOPS），外形小巧，节能高效（7.5W），非常适合机器人、无人机、智能摄像机和便携医疗设备等智能终端设备。 Jatson TX2 与 TX1 相比，内存和 eMMC 提高了一倍，CUDA 架构升级为 Pascal，每瓦性能提高一倍，支持 Jetson TX1
0. 简介
Jetson TX2【1】是基于 NVIDIA Pascal™ 架构的 AI 单模块超级计算机，性能强大（1 TFLOPS），外形小巧，节能高效（7.5W），非常适合机器人、无人机、智能摄像机和便携医疗设备等智能终端设备。
Jatson TX2 与 TX1 相比，内存和 eMMC 提高了一倍，CUDA 架构升级为 Pascal，每瓦性能提高一倍，支持 Jetson TX1 模块的所有功能，支持更大、更深、更复杂的深度神经网络。
TX2 内部结构如下：
1. 开箱
过程细节不展开，板卡上电后来张照片：
2. 刷机
TX2 出厂时，已经自带了 Ubuntu 16.04 系统，可以直接启动。但一般我们会选择刷机，目的是更新到最新的 JetPack L4T，并自动安装最新的驱动、CUDA Toolkit、cuDNN、TensorRT。
刷机注意以下几点：
Host 需要安装 Ubuntu 14.04，至少预留 15 GB 硬盘空间，不要用 root 用户运行 JetPack-${VERSION}.run，我用的是 JetPack-L4T-3.1-linux-x64.run
TX2 需要进入 Recovery Mode，参考随卡自带的说明书步骤
刷机时间大概需要 1~2 小时，会格式化 eMMC，主要备份数据
3. 运行视频目标检测 Demo
刷机成功后，重启 TX2，连接键盘鼠标显示器，就可以跑 Demo 了。
nvidia@tegra-ubuntu:~/tegra_multimedia_api/samples/backend$ ./backend1../../data/Video/sample_outdoor_car_1080p_10fps.h264 H264 --trt-deployfile ../../data/Model/GoogleNet_one_class/GoogleNet_modified_oneClass_halfHD.prototxt --trt-modelfile ../../data/Model/GoogleNet_one_class/GoogleNet_modified_oneClass_halfHD.caffemodel --trt-forcefp320--trt-proc-interval1-fps10
视频截图如下：
4. 运行 TensorRT Benchmark
TensorRT 【3】是 Nvidia GPU 上的深度学习 inference 优化库，可以将训练好的模型通过优化器生成 inference 引擎
将 TX2 设置为 MAXP （最高性能）模式，运行 TensorRT 加速的 GoogLeNet、VGG16 得到处理性能如下：
5. TX2 不支持的 feature
不支持 int8
待发现
参考文献
【1】嵌入式系统开发者套件和模块 | NVIDIA Jetson | NVIDIA
【2】Download and Install JetPack L4T
【3】TensorRT
附录
deviceQuery
nvidia@tegra-ubuntu:~/work/TensorRT/tmp/usr/src/tensorrt$ cd /usr/local/cuda/samples/1_Utilities/deviceQuerynvidia@tegra-ubuntu:/usr/local/cuda/samples/1_Utilities/deviceQuery$ lsdeviceQuery deviceQuery.cpp deviceQuery.o Makefile NsightEclipse.xml readme.txtnvidia@tegra-ubuntu:/usr/local/cuda/samples/1_Utilities/deviceQuery$ ./deviceQuery./deviceQuery Starting... CUDA Device Query (Runtime API) version (CUDART static linking)Detected 1 CUDA Capable device(s)Device 0: "NVIDIA Tegra X2" CUDA Driver Version / Runtime Version 8.0 / 8.0 CUDA Capability Major/Minor version number: 6.2 Total amount of global memory: 7851 MBytes (8232062976 bytes) ( 2) Multiprocessors, (128) CUDA Cores/MP: 256 CUDA Cores GPU Max Clock rate: 1301 MHz (1.30 GHz) Memory Clock rate: 1600 Mhz Memory Bus Width: 128-bit L2CacheSize:524288bytesMaximum TextureDimensionSize(x,y,z)1D=(131072),2D=(131072,65536),3D=(16384,16384,16384) Maximum Layered1D TextureSize, (num) layers1D=(32768),2048layers Maximum Layered2D TextureSize, (num) layers2D=(32768,32768),2048layers Total amountofconstantmemory:65536bytesTotal amountofsharedmemoryperblock:49152bytesTotalnumberofregisters available perblock:32768Warpsize:32Maximumnumberofthreads per multiprocessor:2048Maximumnumberofthreads perblock:1024Maxdimensionsizeofathreadblock(x,y,z): (1024,1024,64)Maxdimensionsizeofa gridsize(x,y,z): (2147483647,65535,65535) Maximummemorypitch:2147483647bytesTexture alignment:512bytesConcurrentcopyandkernel execution: Yeswith1copyengine(s) Runtimelimitonkernels:NoIntegrated GPU sharing HostMemory: Yes Support host page-lockedmemorymapping: Yes Alignment requirementforSurfaces: Yes Device has ECC support: Disabled Device supports Unified Addressing (UVA): Yes Device PCIDomainID/ BusID/ locationID:0/0/0ComputeMode: deviceQuery, CUDA Driver = CUDART, CUDA DriverVersion=8.0, CUDA RuntimeVersion=8.0, NumDevs =1, Device0 = NVIDIA Tegra X2Result= PASS
内存带宽测试
nvidia@tegra-ubuntu:/usr/local/cuda/samples/1_Utilities/bandwidthTest$ ./bandwidthTest[CUDABandwidthTest] - Starting...Runningon... Device0: NVIDIA Tegra X2 Quick Mode Host to DeviceBandwidth,1Device(s) PINNED Memory Transfers Transfer Size (Bytes)Bandwidth(MB/s)3355443220215.8Device to HostBandwidth,1Device(s) PINNED Memory Transfers Transfer Size (Bytes)Bandwidth(MB/s)3355443220182.2Device to DeviceBandwidth,1Device(s) PINNED Memory Transfers Transfer Size (Bytes)Bandwidth(MB/s)3355443235742.8Result= PASSNOTE:The CUDA Samples are not meant for performance measurements. Results may vary when GPUBoostis enabled.
GEMM 测试
nvidia@tegra-ubuntu:/usr/local/cuda/samples/7_CUDALibraries/batchCUBLAS$ ./batchCUBLAS -m1024 -n1024 -k1024batchCUBLAS Starting...GPU Device 0: "NVIDIA Tegra X2" with compute capability 6.2 ==== Running single kernels ====Testingsgemm#### args: ta=0 tb=0 m=1024 n=1024 k=1024 alpha = (0xbf800000,-1) beta= (0x40000000, 2)#### args: lda=1024 ldb=1024 ldc=1024^^^^ elapsed = 0.00372291 sec GFLOPS=576.83@@@@ sgemm test OKTestingdgemm#### args: ta=0 tb=0 m=1024 n=1024 k=1024 alpha = (0x0000000000000000, 0) beta= (0x0000000000000000, 0)#### args: lda=1024 ldb=1024 ldc=1024^^^^ elapsed = 0.10940003 sec GFLOPS=19.6296@@@@ dgemm test OK ==== Running N=10 without streams ====Testingsgemm#### args: ta=0 tb=0 m=1024 n=1024 k=1024 alpha = (0xbf800000,-1) beta= (0x00000000, 0)#### args: lda=1024 ldb=1024 ldc=1024^^^^ elapsed = 0.03462315 sec GFLOPS=620.245@@@@ sgemm test OKTestingdgemm#### args: ta=0 tb=0 m=1024 n=1024 k=1024 alpha = (0xbff0000000000000,-1) beta= (0x0000000000000000, 0)#### args: lda=1024 ldb=1024 ldc=1024^^^^ elapsed = 1.09212208 sec GFLOPS=19.6634@@@@ dgemm test OK ==== Running N=10 with streams ====Testingsgemm#### args: ta=0 tb=0 m=1024 n=1024 k=1024 alpha = (0x40000000, 2) beta= (0x40000000, 2)#### args: lda=1024 ldb=1024 ldc=1024^^^^ elapsed = 0.03504515 sec GFLOPS=612.776@@@@ sgemm test OKTestingdgemm#### args: ta=0 tb=0 m=1024 n=1024 k=1024 alpha = (0xbff0000000000000,-1) beta= (0x0000000000000000, 0)#### args: lda=1024 ldb=1024 ldc=1024^^^^ elapsed = 1.09177494 sec GFLOPS=19.6697@@@@ dgemm test OK ==== Running N=10 batched ====Testingsgemm#### args: ta=0 tb=0 m=1024 n=1024 k=1024 alpha = (0x3f800000, 1) beta= (0xbf800000,-1)#### args: lda=1024 ldb=1024 ldc=1024^^^^ elapsed = 0.03766394 sec GFLOPS=570.17@@@@ sgemm test OKTestingdgemm#### args: ta=0 tb=0 m=1024 n=1024 k=1024 alpha = (0xbff0000000000000,-1) beta= (0x4000000000000000, 2)#### args: lda=1024 ldb=1024 ldc=1024^^^^ elapsed = 1.09389901 sec GFLOPS=19.6315@@@@ dgemm test OKTestSummary0 error(s)