1.2ver gpu can;t workers

[慈馹吃]

An error occured while executing forward of function Affine (nn.AffineCuda) in network Main

Failed `status == CUBLAS_STATUS_SUCCESS`: ARCH_MISMATCH

1.1 1.0 was pass

gpu 780M

[Yoshiyuki Kobayashi]

Neural Network Console v1.20 uses CUDA 9.1 instead of CUDA 8.0 used before.
Could you try updating the driver of the GPU?

http://www.nvidia.com/Download/index.aspx

[vvinewyork]

Installed "the latest" CUDA, which is 9.2, as was recommended in the setup instructions.

Installed NN Console 1.2.

Had exactly the same error message, in the same function.

Uninstalled CUDA 9.2, installed CUDA 9.1. The same message again. Had to re-install ver.1.1 of NN Console.

NN Console v. 1.1 works fine with CUDA 9.1; v.1.2 - doesn't.

GPU/drivers information is included below:

NVIDIA System Information report created on: 06/25/2018 00:08:24

[Display]
Operating System: Windows 7 Ultimate, 64-bit (Service Pack 1)
DirectX version: 11.0
GPU processor:  Quadro K6000
Driver version:  388.19
Direct3D API version: 11
Direct3D feature level: 11_0
CUDA Cores:  2880
Core clock:  797 MHz
Memory data rate: 6008 MHz
Memory interface: 384-bit
Memory bandwidth: 288.38 GB/s
Total available graphics memory: 28388 MB
Dedicated video memory: 12288 MB GDDR5
System video memory: 0 MB
Shared system memory: 16100 MB
Video BIOS version: 80.80.19.00.01
IRQ:   Not used
Bus:   PCI Express x16 Gen3
Error Correction Code (ECC): Off
Device Id:  10DE 103A 103610DE
Part Number:  2081 0500

[Components]

nvui.dll  8.17.13.8819  NVIDIA User Experience Driver Component
nvxdplcy.dll  8.17.13.8819  NVIDIA User Experience Driver Component
nvxdbat.dll  8.17.13.8819  NVIDIA User Experience Driver Component
nvxdapix.dll  8.17.13.8819  NVIDIA User Experience Driver Component
NVCPL.DLL  8.17.13.8819  NVIDIA User Experience Driver Component
nvCplUIR.dll  8.1.940.0  NVIDIA Control Panel
nvCplUI.exe  8.1.940.0  NVIDIA Control Panel
nvWSSR.dll  23.21.13.8819  NVIDIA Workstation Server
nvWSS.dll  23.21.13.8819  NVIDIA Workstation Server
nvViTvSR.dll  23.21.13.8819  NVIDIA Video Server
nvViTvS.dll  23.21.13.8819  NVIDIA Video Server
NVSTVIEW.EXE  7.17.13.8819  NVIDIA 3D Vision Photo Viewer
NVSTTEST.EXE  7.17.13.8819  NVIDIA 3D Vision Test Application
NVSTRES.DLL  7.17.13.8819  NVIDIA 3D Vision Module
nvDispSR.dll  23.21.13.8819  NVIDIA Display Server
NVMCTRAY.DLL  23.21.13.8819  NVIDIA Media Center Library
nvDispS.dll  23.21.13.8819  NVIDIA Display Server
PhysX   09.17.0524  NVIDIA PhysX
NVCUDA.DLL  23.21.13.8819  NVIDIA CUDA 9.1.83 driver
nvGameSR.dll  23.21.13.8819  NVIDIA 3D Settings Server
nvGameS.dll  23.21.13.8819  NVIDIA 3D Settings Server

[Yoshiyuki Kobayashi]

CUDA runtime libraries are included in the Neural Network Console.
So, basically you only need to update a GPU driver to the latest to use NNC v1.2.

However, it seems from the information you posted that NNC v1.2 does not work on GPU with Kepler architecture.
Could you please try the following procedure?

1. Install CUDA 8.0 and cuDNN 7.0.
2. From command prompt:

SET PATH=%PATH%;C:\(NNC v1.2 installation dir)\libs\Miniconda3;C:\(NNC v1.2 installation dir)\libs\Miniconda3\Scripts
pip uninstall nnabla-ext-cuda
pip install nnabla-ext-cuda80==1.0.0

[vvinewyork]

Thank you for your reply.

Ok, time permitting I will try it again. However, all these install/uninstall games take a lot of time, andpresent difference between NNC versions is not big enough to go al the way back to CUDA 8.0.

CUDA installer automatically installs all necessary drivers, and it does work for all supported architectures, including Kepler.

However, if I try to install "the latest" GPU driver manually, it is very easy to create a mess by installing the wrong one.

Also, in my humble opinion, you should re-evaluate your NNC deployment strategy. If your application has CUDA dependency, it is much better to defer the installation of all libraries to the original vendor (NVIDIA in this instance).

So, instead of distributing CUDA runtime with your application, the application should detect and use the version already installed on the customer's machine.