Announcement

Collapse
No announcement yet.

Five-Way Linux OS Comparison On Amazon's ARM Graviton CPU

Collapse
X
 
  • Filter
  • Time
  • Show
Clear All
new posts

  • Five-Way Linux OS Comparison On Amazon's ARM Graviton CPU

    Phoronix: Five-Way Linux OS Comparison On Amazon's ARM Graviton CPU

    Last month Amazon rolled out their "Graviton" ARM processors in the Elastic Compute Cloud. Those first-generation Graviton ARMv8 processors are based on the ARM Cortex-A72 cores and designed to offer better pricing than traditional x86_64 EC2 instances. However, our initial testing of the Amazon Graviton EC2 "A1" instances didn't reveal significant performance-per-dollar benefits for these new instances. In this second round of Graviton CPU benchmarking we are seeing what is the fastest of five of the leading ARM Linux distributions.

    Phoronix, Linux Hardware Reviews, Linux hardware benchmarks, Linux server benchmarks, Linux benchmarking, Desktop Linux, Linux performance, Open Source graphics, Linux How To, Ubuntu benchmarks, Ubuntu hardware, Phoronix Test Suite

  • #2
    hello michael and phoronix members

    For deep learning benchmark on ARM devices, you may interest in the ncnn opensource project.
    ncnn is a high-performance neural network inference framework optimized for the mobile platform.
    It is a heavily hand-optimized and widely-used NN inference framework. It makes extensive use of NEON assembly instructions.

    ncnn is a high-performance neural network inference framework optimized for the mobile platform - Tencent/ncnn


    Many companies and individual developers use caffe/mxnet/pytorch/tensorflow/... for training and use ncnn for deploying on arm devices.

    https://github.com/Tencent/ncnn/wiki...th-ncnn-inside

    There is a tool bundled for easy benchmarking, and results for some CNN models posted

    ncnn is a high-performance neural network inference framework optimized for the mobile platform - Tencent/ncnn



    best wishes

    Comment


    • #3
      Michael it seems a nice test case

      Comment


      • #4
        Originally posted by nihui View Post
        hello michael and phoronix members

        For deep learning benchmark on ARM devices, you may interest in the ncnn opensource project.
        ncnn is a high-performance neural network inference framework optimized for the mobile platform.
        It is a heavily hand-optimized and widely-used NN inference framework. It makes extensive use of NEON assembly instructions.

        ncnn is a high-performance neural network inference framework optimized for the mobile platform - Tencent/ncnn


        Many companies and individual developers use caffe/mxnet/pytorch/tensorflow/... for training and use ncnn for deploying on arm devices.

        https://github.com/Tencent/ncnn/wiki...th-ncnn-inside

        There is a tool bundled for easy benchmarking, and results for some CNN models posted

        ncnn is a high-performance neural network inference framework optimized for the mobile platform - Tencent/ncnn



        best wishes
        With not being familiar with ncnn, happen to have a reference to a complete benchmark script for it?
        Michael Larabel
        https://www.michaellarabel.com/

        Comment


        • #5
          Originally posted by Michael View Post

          With not being familiar with ncnn, happen to have a reference to a complete benchmark script for it?
          https://github.com/Tencent/ncnn/tree/master/benchmark provides the instructions. Basically, you need to build the benchncnn.cpp file and then run the executable. It enumerates the current directory for all param files for benchmarks to run.

          The output looks well structured so it's definitely possible to put that into PTS.

          Comment


          • #6
            Originally posted by Michael View Post

            With not being familiar with ncnn, happen to have a reference to a complete benchmark script for it?
            hello Michael

            you can follow the how-to-build wiki for bulding the ncnn library
            ncnn is a high-performance neural network inference framework optimized for the mobile platform - Tencent/ncnn


            and for the benchmark tool, that's the benchncnn executable, it will print the inference time consumed for each CNN model
            https://github.com/Tencent/ncnn/blob...mark/README.md

            best wishes
            Last edited by nihui; 23 December 2018, 10:20 PM.

            Comment


            • #7
              Michael hi, we have recently implement vulkan compute support into ncnn library, which brings universal cross-platform gpu inference on desktop and mobile. At the moment, some basic CNN network models already work fine, and we are working hard on optimizing the performance

              new benchmark or opensource software news would be appreciated. thanks

              Comment


              • #8
                Originally posted by nihui View Post
                Michael hi, we have recently implement vulkan compute support into ncnn library, which brings universal cross-platform gpu inference on desktop and mobile. At the moment, some basic CNN network models already work fine, and we are working hard on optimizing the performance

                new benchmark or opensource software news would be appreciated. thanks
                Got a build and run script for it yet for benchmarking to make it easy and automated to complete?
                Michael Larabel
                https://www.michaellarabel.com/

                Comment


                • #9
                  Originally posted by Michael View Post

                  Got a build and run script for it yet for benchmarking to make it easy and automated to complete?

                  1. download and setup vulkan sdk
                  2. clone ncnn source
                  3. build ncnn with vulkan option on
                  4. run benchncnn tool
                  5. speed info will be output on stderr

                  ncnn is a high-performance neural network inference framework optimized for the mobile platform - Tencent/ncnn


                  sample script

                  Code:
                  $ wget https://sdk.lunarg.com/sdk/download/1.1.92.1/linux/vulkansdk-linux-x86_64-1.1.92.1.tar.gz?Human=true -O vulkansdk-linux-x86_64-1.1.92.1.tar.gz
                  $ tar -xf vulkansdk-linux-x86_64-1.1.92.1.tar.gz
                  $ export VULKAN_SDK=`pwd`/1.1.92.1/x86_64
                  
                  $ git clone https://github.com/Tencent/ncnn.git
                  $ cd ncnn
                  $ mkdir -p build && cd build
                  $ cmake -DNCNN_VULKAN=ON ..
                  $ make -j4
                  
                  $ cp ../benchmark/*.param benchmark/
                  $ cd benchmark
                  
                  ### run on host cpu (4 threads)
                  $ ./benchncnn 10 4 0 -1
                  
                  ### run on host gpu-0
                  $ ./benchncnn 10 4 0 0

                  Comment


                  • #10
                    Originally posted by nihui View Post


                    1. download and setup vulkan sdk
                    2. clone ncnn source
                    3. build ncnn with vulkan option on
                    4. run benchncnn tool
                    5. speed info will be output on stderr

                    ncnn is a high-performance neural network inference framework optimized for the mobile platform - Tencent/ncnn


                    sample script

                    Code:
                    $ wget https://sdk.lunarg.com/sdk/download/1.1.92.1/linux/vulkansdk-linux-x86_64-1.1.92.1.tar.gz?Human=true -O vulkansdk-linux-x86_64-1.1.92.1.tar.gz
                    $ tar -xf vulkansdk-linux-x86_64-1.1.92.1.tar.gz
                    $ export VULKAN_SDK=`pwd`/1.1.92.1/x86_64
                    
                    $ git clone https://github.com/Tencent/ncnn.git
                    $ cd ncnn
                    $ mkdir -p build && cd build
                    $ cmake -DNCNN_VULKAN=ON ..
                    $ make -j4
                    
                    $ cp ../benchmark/*.param benchmark/
                    $ cd benchmark
                    
                    ### run on host cpu (4 threads)
                    $ ./benchncnn 10 4 0 -1
                    
                    ### run on host gpu-0
                    $ ./benchncnn 10 4 0 0
                    Thanks! Will try it out this weekend.
                    Michael Larabel
                    https://www.michaellarabel.com/

                    Comment

                    Working...
                    X