[SOLVED] Two Radeon HD 5870 in one computer, but only one working?

[kefir]

Hi,

I have a computer with 2 ATI Radeon HD 5870 cards. I have got things up and running, but OpenCL can only see one GPU, not both.

I've installed Ubuntu 11.04 Natty Narwhal, and I've installed the fglrx driver (8.840) and the AMD APP SDK (2.4), built pyopencl, and have successfully run poclbm on one GPU, and also on the CPU (though jgarzik's cpuminer was a lot faster on the CPU using sse2_64).

Anyone have any idea what I need to do to get both cards mining?

lspci shows both cards:

Code:

04:00.0 VGA compatible controller: ATI Technologies Inc Radeon HD 5870 (Cypress)
09:00.0 VGA compatible controller: ATI Technologies Inc Radeon HD 5870 (Cypress)

I've configured both cards in xorg.conf:

Code:

Section "Device"
        Identifier "radeon0"
        Driver "fglrx"
        BusID "PCI:9:0:0"
EndSection

Section "Device"
        Identifier "radeon1"
        Driver "fglrx"
        BusID "PCI:4:0:0"
EndSection

and X starts up fine, but clinfo (bundled with AMD APP SDK) just shows the GPU and the CPU:

Code:

Number of platforms:				 1
  Platform Profile:				 FULL_PROFILE
  Platform Version:				 OpenCL 1.1 AMD-APP-SDK-v2.4 (595.10)
  Platform Name:				 AMD Accelerated Parallel Processing
  Platform Vendor:				 Advanced Micro Devices, Inc.
  Platform Extensions:				 cl_khr_icd cl_amd_event_callback cl_amd_offline_devices


  Platform Name:				 AMD Accelerated Parallel Processing
Number of devices:				 2
  Device Type:					 CL_DEVICE_TYPE_GPU
  Device ID:					 4098
  Max compute units:				 20
  Max work items dimensions:			 3
    Max work items[0]:				 256
    Max work items[1]:				 256
    Max work items[2]:				 256
  Max work group size:				 256
  Preferred vector width char:			 16
  Preferred vector width short:			 8
  Preferred vector width int:			 4
  Preferred vector width long:			 2
  Preferred vector width float:			 4
  Preferred vector width double:		 0
  Native vector width char:			 16
  Native vector width short:			 8
  Native vector width int:			 4
  Native vector width long:			 2
  Native vector width float:			 4
  Native vector width double:			 0
  Max clock frequency:				 850Mhz
  Address bits:					 32
  Max memory allocation:			 134217728
  Image support:				 Yes
  Max number of images read arguments:		 128
  Max number of images write arguments:		 8
  Max image 2D width:				 8192
  Max image 2D height:				 8192
  Max image 3D width:				 2048
  Max image 3D height:				 2048
  Max image 3D depth:				 2048
  Max samplers within kernel:			 16
  Max size of kernel argument:			 1024
  Alignment (bits) of base address:		 32768
  Minimum alignment (bytes) for any datatype:	 128
  Single precision floating point capability
    Denorms:					 No
    Quiet NaNs:					 Yes
    Round to nearest even:			 Yes
    Round to zero:				 Yes
    Round to +ve and infinity:			 Yes
    IEEE754-2008 fused multiply-add:		 Yes
  Cache type:					 None
  Cache line size:				 0
  Cache size:					 0
  Global memory size:				 536870912
  Constant buffer size:				 65536
  Max number of constant args:			 8
  Local memory type:				 Scratchpad
  Local memory size:				 32768
  Kernel Preferred work group size multiple:	 64
  Error correction support:			 0
  Unified memory for Host and Device:		 0
  Profiling timer resolution:			 1
  Device endianess:				 Little
  Available:					 Yes
  Compiler available:				 Yes
  Execution capabilities:				 
    Execute OpenCL kernels:			 Yes
    Execute native function:			 No
  Queue properties:				 
    Out-of-Order:				 No
    Profiling :					 Yes
  Platform ID:					 0x7f1b34e26800
  Name:						 Cypress
  Vendor:					 Advanced Micro Devices, Inc.
  Driver version:				 CAL 1.4.1353
  Profile:					 FULL_PROFILE
  Version:					 OpenCL 1.1 AMD-APP-SDK-v2.4 (595.10)
  Extensions:					 cl_amd_fp64 cl_khr_global_int32_base_atomics cl_khr_global_int32_extended_atomics cl_khr_local_int32_base_atomics cl_khr_local_int32_extended_atomics cl_khr_3d_image_writes cl_khr_byte_addressable_store cl_khr_gl_sharing cl_amd_device_attribute_query cl_amd_printf cl_amd_media_ops cl_amd_popcnt 


  Device Type:					 CL_DEVICE_TYPE_CPU
  Device ID:					 4098
  Max compute units:				 6
  Max work items dimensions:			 3
    Max work items[0]:				 1024
    Max work items[1]:				 1024
    Max work items[2]:				 1024
  Max work group size:				 1024
  Preferred vector width char:			 16
  Preferred vector width short:			 8
  Preferred vector width int:			 4
  Preferred vector width long:			 2
  Preferred vector width float:			 4
  Preferred vector width double:		 0
  Native vector width char:			 16
  Native vector width short:			 8
  Native vector width int:			 4
  Native vector width long:			 2
  Native vector width float:			 4
  Native vector width double:			 0
  Max clock frequency:				 800Mhz
  Address bits:					 64
  Max memory allocation:			 2147483648
  Image support:				 Yes
  Max number of images read arguments:		 128
  Max number of images write arguments:		 8
  Max image 2D width:				 8192
  Max image 2D height:				 8192
  Max image 3D width:				 2048
  Max image 3D height:				 2048
  Max image 3D depth:				 2048
  Max samplers within kernel:			 16
  Max size of kernel argument:			 4096
  Alignment (bits) of base address:		 1024
  Minimum alignment (bytes) for any datatype:	 128
  Single precision floating point capability
    Denorms:					 Yes
    Quiet NaNs:					 Yes
    Round to nearest even:			 Yes
    Round to zero:				 Yes
    Round to +ve and infinity:			 Yes
    IEEE754-2008 fused multiply-add:		 No
  Cache type:					 Read/Write
  Cache line size:				 64
  Cache size:					 65536
  Global memory size:				 8388317184
  Constant buffer size:				 65536
  Max number of constant args:			 8
  Local memory type:				 Global
  Local memory size:				 32768
  Kernel Preferred work group size multiple:	 1
  Error correction support:			 0
  Unified memory for Host and Device:		 1
  Profiling timer resolution:			 1
  Device endianess:				 Little
  Available:					 Yes
  Compiler available:				 Yes
  Execution capabilities:				 
    Execute OpenCL kernels:			 Yes
    Execute native function:			 Yes
  Queue properties:				 
    Out-of-Order:				 No
    Profiling :					 Yes
  Platform ID:					 0x7f1b34e26800
  Name:						 AMD Phenom(tm) II X6 1055T Processor
  Vendor:					 AuthenticAMD
  Driver version:				 2.0
  Profile:					 FULL_PROFILE
  Version:					 OpenCL 1.1 AMD-APP-SDK-v2.4 (595.10)
  Extensions:					 cl_khr_fp64 cl_amd_fp64 cl_khr_global_int32_base_atomics cl_khr_global_int32_extended_atomics cl_khr_local_int32_base_atomics cl_khr_local_int32_extended_atomics cl_khr_int64_base_atomics cl_khr_int64_extended_atomics cl_khr_byte_addressable_store cl_khr_gl_sharing cl_ext_device_fission cl_amd_device_attribute_query cl_amd_vec3 cl_amd_media_ops cl_amd_popcnt cl_amd_printf 

[1714813925]

1714813925

[1714813925]

1714813925

[anodyne]

Not sure if it applies to Linux, but the Radeons have a power saving state that disables them if no screen is connected. You can start mining on one, and then plug your screen into the other and start a miner there as a temporary workaround.

Less temporary would be a dummy plug. A forum search will show you how to make one out of the DVI-VGA adapter and some resistors.

[Folax]

I had 2 5870s in my WinXP box working both (only a screen attached to 1 card), then I had to reboot, and afterwards only 1 card was active, the 2nd one was disabled by ATI CCC.

I installed LinuxCoin and both cards work fine now, with still only 1 screen attached (to 1 card).

I never used a dummy plug or 2 screens, I really don't know why I don't need to though.

Try LinuxCoin on a USB drive and see if your cards work there I'd suggest.

[kefir]

Brilliant, that worked! Thanks! Well, it wasn't enough to just connect a second screen, I actually had to modify my xorg.conf ServerLayout so that both screens are active to get it working.

Here's my current xorg.conf. I'm not sure all of this is necessary, but at least it works for me now. Getting about 340 Mhash/s on each card, not overclocked.

Code:

Section "Module"
        Load    "dri"
        Load    "glx"
EndSection

Section "ServerLayout"
        Identifier "Standard config"
        Screen 0 "screen0" 0 0
        Screen "screen1" LeftOf "screen0"
        InputDevice    "Mouse0" "CorePointer"
        InputDevice    "Keyboard0" "CoreKeyboard"
EndSection

Section "InputDevice"
        Identifier  "Keyboard0"
        Driver      "kbd"
EndSection

Section "InputDevice"
        Identifier  "Mouse0"
        Driver      "mouse"
        Option      "Protocol" "auto"
        Option      "Device" "/dev/input/mice"
        Option      "ZAxisMapping" "4 5 6 7"
EndSection

Section "Monitor"
        Identifier "monitor0"
        Option "VendorName" "Fujitsu-Siemens"
        Option "ModelName" "SCALEOVIEW T19-2"
EndSection

Section "Monitor"
        Identifier "monitor1"
        Option "VendorName" "Vendor"
        Option "ModelName" "Model"
EndSection

Section "Device"
        Identifier "radeon0"
        Driver "fglrx"
        BusID "PCI:9:0:0"
EndSection

Section "Device"
        Identifier "radeon1"
        Driver "fglrx"
        BusID "PCI:4:0:0"
EndSection

Section "Screen"
        Identifier "screen0"
        Device "radeon0"
        Monitor "monitor0"
        DefaultDepth 24
        SubSection "Display"
                Viewport 0 0
                Depth 24
        EndSubSection
EndSection

Section "Screen"
        Identifier "screen1"
        Device "radeon1"
        Monitor "monitor1"
        DefaultDepth 24
        SubSection "Display"
                Viewport 0 0
                Depth 24
        EndSubSection
EndSection

Folax, thanks, I'll check out LinuxCoin. Would be great to avoid multiple screens and dummy plugs.

[anodyne]

The disabling seems to be on the driver level. On Windows there's a registry setting for ULPS that works according to some, but apparently not for everyone. Maybe someone managed to put the proper setting into the default Linuxcoin settings?

I have also had to extend the desktop on Windows to get my miner to recognize secondary, dummy plugged GPUs. Seems like there's always one new problem waiting around the corner...

[Folax]

To me it seems that with Windows 'luck' is the factor that decides if it will work or not, on Linux you are depending on 'a problem that can be solved'.