dividiti / ck-caffe-nvidia-gtx1080

CK-Caffe public benchmarking data on NVIDIA GTX1080

cknowledge.org/ai: Crowdsourcing benchmarking and optimisation of AI

[PUBLIC] Benchmarking Caffe on NVIDIA GTX 1080

NB: The Caffe experimental results are released with approval from General Motors.

The Jupyter notebook (view on github.com; view on nbviewer.jupyter.org) in this Collective Knowledge repository analyses the performance (execution time, memory consumption):

• on dividiti's velociti Hewlett-Packard Z640 Workstation (G1X62EA):
  • Intel(R) Xeon(R) CPU E5-2650 v3:
    • 10 cores, 20 threads;
    • Base clock 2300 MHz, turbo clock 3000 MHz;
    • Max power consumption 105 Watt;
    • Max memory bandwidth 68 GB/s;
    • RAM memory 32 GB DDR4;

$ uname -a
Linux velociti 4.4.0-45-generic #66-Ubuntu SMP Wed Oct 19 14:12:37 UTC 2016 x86_64 x86_64 x86_64 GNU/Linux
$ cat /etc/lsb-release
DISTRIB_ID=Ubuntu
DISTRIB_RELEASE=16.04
DISTRIB_CODENAME=xenial
DISTRIB_DESCRIPTION="Ubuntu 16.04.1 LTS"

• NVIDIA GeForce GTX 1080 "Founders Edition":

  • Pascal architecture;
  • 2560 CUDA cores;
  • Base clock 1607 MHz, boost clock 1733 MHz;
  • Max power consumption 180 Watt;
  • RAM memory 8 GB GDDR5X;
  • Max memory bandwidth 320 GB/s;
  • GPU Driver 367.57 [10/Oct/2016];
  • CUDA Toolkit 8.0.44 [xx/Sep/2016].

• using 14 Caffe libraries:

  • [tag] Branch (revision hash, date): math libraries.
  • [cpu] Master (4ba654f, 5/Oct/2016): with OpenBLAS 0.2.19;
  • [cuda] Master (4ba654f, 5/Oct/2016): with cuBLAS (part of CUDA Toolkit 8.0.44);
  • [cudnn] Master (4ba654f, 5/Oct/2016): with cuDNN 5.1;
  • [nvidia-cuda] NVIDIA v0.15 (1024d34, 17/Nov/2016): with cuBLAS (part of CUDA Toolkit 8.0.44);
  • [nvidia-cudnn] NVIDIA v0.15 (1024d34, 17/Nov/2016): with cuDNN 5.1;
  • [nvidia-fp16-cuda] NVIDIA experimental/fp16 (fca1cf4, 11/Jul/2016): with cuBLAS (part of CUDA Toolkit 8.0.44);
  • [nvidia-fp16-cudnn] NVIDIA experimental/fp16 (fca1cf4, 11/Jul/2016): with cuDNN 5.1;
  • [clblas] OpenCL (9abafdc, 7/Oct/2016): with ViennaCL 1.7.1 and clBLAS 2.10;
  • [clblast] OpenCL (9abafdc, 7/Oct/2016): with ViennaCL 1.7.1 and CLBlast 0.9.0;
  • [viennacl] OpenCL (9abafdc, 7/Oct/2016): with ViennaCL 1.7.1 only;
  • [libdnn-cuda] OpenCL (cfaaae1, 25/Oct/2016): with libDNN and cuBLAS;
  • [libdnn-clblas] OpenCL (cfaaae1, 25/Oct/2016): with libDNN, ViennaCL 1.7.1 and clBLAS 2.10;
  • [libdnn-clblast] OpenCL (cfaaae1, 25/Oct/2016): with libDNN, ViennaCL 1.7.1 and CLBlast 0.9.0;
  • [libdnn-viennacl] OpenCL (cfaaae1, 25/Oct/2016): with libDNN and ViennaCL 1.7.1.

• using 4 CNN models:

  • GoogleNet;
  • AlexNet;
  • SqueezeNet 1.0;
  • SqueezeNet 1.1;

• with the batch size varying from 2 to 16 with step 2.