IBM Support

NVidia CUDA Toolkit bandwidth test fails under Microsoft Windows Server 2008 R2 - IBM iDataPlex dx360 M3 Server (6391)

Troubleshooting


Problem

The CUDA bandwidth test that is included in the NVIDIA CUDA Toolkit and SDK package, available from nvidia.com, fails when it is run on the iDataPlex dx360 M3 Server, Type 6391, single-node 2U chassis with the three-slot PCIe riser installed and populated with two (2) Tesla M2050 adapters or two (2) Tesla M2070/M2070Q adapters. Below are the steps to re-create this issue: 1. Configure two (2) Tesla M2050 adapters or Tesla two (2) M2070/M2070Q adapters in the iDataPlex dx360 M3 2U system with Windows Server 2008 R2. 2. Install NVIDIA Win2K8R2 driver, the NVIDIA 'cudatoolkit_3.x' and the NVIDIA 'gpucomputingsdk_3.x'. 3. Run the "bandwidthtest" program to verify configuration of the hardware and software. The test fails and reports the followingmessage: cudaSafeCall() Runtime API error : no CUDA-capable device is available.

Resolving The Problem

Source

RETAIN tip: H197859

Symptom

The CUDA bandwidth test that is included in the NVIDIA CUDA Toolkit and SDK package, available from nvidia.com, fails when it is run on the iDataPlex dx360 M3 Server, Type 6391, single-node 2U chassis with the three-slot PCIe riser installed and populated with two (2) Tesla M2050 adapters or two (2) Tesla M2070/M2070Q adapters.

Below are the steps to re-create this issue:

  1. Configure two (2) Tesla M2050 adapters or Tesla two (2) M2070/M2070Q adapters in the iDataPlex dx360 M3 2U system with Windows Server 2008 R2.

  2. Install NVIDIA Win2K8R2 driver, the NVIDIA 'cudatoolkit_3.x' and the NVIDIA 'gpucomputingsdk_3.x'.

  3. Run the "bandwidthtest" program to verify configuration of the hardware and software.

The test fails and reports the following message:

  cudaSafeCall() Runtime API error : no CUDA-capable device is available.

Affected configurations

The system may be any of the following IBM servers:

  • iDataPlex dx360 M3 Server, type 6391, any model

The system is configured with one or more of the following IBM Options:

  • NVIDIA Tesla M2050 GPU Adapter, Assembly 43V5926, CRU 43V5925
  • NVIDIA Tesla M2070 GPU Adapter, Assembly 43V5936 , CRU 43V5935
  • NVIDIA Tesla M2070Q GPU Adapter, Assembly 43V5944, CRU 43V5943

This tip is not software specific.

  • The CUDA Bandwidth Test utility is affected.
  • The NVIDIA W2K8R2 device driver for the Tesla cards is affected.
  • Tesla cards

Workaround

Enter the registry edit utility on the iDataPlex dx360 M3 and for each device listed under the registry path:

  [HKEY_LOCAL_MACHINE\SYSTEM\CurrentControlSet\Control\Class\{4D36E968-E325-11CE-BFC1-08002BE10318}\

identified by a four0digit number, add the following reg key:

"AdapterType"=dword:00000002

For example:

  [HKEY_LOCAL_MACHINE\SYSTEM\CurrentControlSet\Control\Class\{4D36E968-E325-11CE-BFC1-08002BE10318}\0000]
"AdapterType"=dword:00000002
[HKEY_LOCAL_MACHINE\SYSTEM\CurrentControlSet\Control\Class\{4D36E968-E325-11CE-BFC1-08002BE10318}\0001]
"AdapterType"=dword:00000002

Add this key for all devices 0000 - nnnn.

Additional information

The Workaround identifies all of the CUDA-capable devices to the Operating System (OS) and the CUDA bandwidth test.

 

Document Location

Worldwide

Operating System

System x:Operating system independent / None

[{"Type":"HW","Business Unit":{"code":"BU016","label":"Multiple Vendor Support"},"Product":{"code":"HWXF0","label":"System x->System x iDataPlex dx360 M3 server"},"Platform":[{"code":"PF025","label":"Platform Independent"}],"Line of Business":{"code":"","label":""}}]

Document Information

Modified date:
30 January 2019

UID

ibm1MIGR-5086246