Reversed GPU 0/GPU 1 designations?

[JimF]

I am folding on two GT 240s. The one in the top PCIe slot (Asus P5Q Pro motherboard) is connected to the display, and the one in the bottom PCIe slot is left unconnected. In the past I have always assumed that the top (display) card was GPU 0, and the bottom card GPU 1, which is what worked for me in setting up FAH Version 6, though that is something of trial-by-error with multiple cards anyway, so don't quote me on that.

When I set up Version 7.1.24, I left the "GPU ID" entry at -1 in the folding slot configuration menu, so that the cards could sort themselves out automatically. But then I found out upon disabling folding for each slot that GPU 0, which should presumably be the top (display) card, was actually the bottom card and vice-versa (as confirmed by the "GPU Load" reading in GPU-Z). I am using the latest Nvidia drivers (275.33) under Win7 64-bit.

The card designations seem to stable, just in a reversed order to what I had assumed before. But to be sure that they stay that way, I manually set the top card to be GPU ID 1. (The bottom card insists on the "GPU ID -1" setting, and will revert to that even if I set it to "GPU ID 0" when I stop/restart the FAHContol.)

Is this reverse designation a problem or not?

[MtM]

There is a difference between gpu-index and -gpu.

Gpu-index corresponds to the 'system' tab, -gpu x can be added as extra-core-argument.

If you want to reverse the designation, manually set gpu-index for the top card to 0 and the bottom card to 1.

Edit: gpu-index has to be set using configure -> slots -> double click on the slot -> extra slot options/add, name=gpu-index, value=0 or 1 depending on which slot offcourse

[JimF]

OK, thanks. But ID 0 shows up as "GPU: 0", and ID 1 shows as "GPU: 1" in the Folding slots description, so you would never guess otherwise.

[codysluder]

I am folding on two GT 240s. The one in the top PCIe slot (Asus P5Q Pro motherboard) is connected to the display, and the one in the bottom PCIe slot is left unconnected. In the past I have always assumed that the top (display) card was GPU 0, and the bottom card GPU 1, which is what worked for me in setting up FAH Version 6, though that is something of trial-by-error with multiple cards anyway, so don't quote me on that.

I don't think there is a universal way of numbering the GPUs. I think your assumption that V7 does it the same way V6 did is the problem, not what they're called.

Assume for a moment that you have never run V6. Is the numbering system used by V7 logical and is V7 internally consistent?

How does the driver number them? More like V6 or more like V7?

[JimF]

The convention I have always heard around here is that GPU 0 is the display card, though I don't know how that is determined.

For the display card (top slot), the Location information is PCI bus 1, device 0, function 0 and Physical Device Object name is \Device\NTPNP_PCI0019.
For the non-display card (bottom slot), the location information is PCI bus 2, device 0, function 0 and the Physical Device Object name is \Device\NTPNP_PCI0021

But when I disabled the non-display driver to detect which one it was, I found that the FAH Client Control Folding Slots "Description" now shows the display card (the only one running) as ID 0, gpu:0. So it has switched around from its previous designation. Then when I enabled both drivers and restarted FAHControl, the display card was again ID 1, gpu:1. That may make perfect sense, but it shows that the gpu designations are not at all fixed, and I think that may be different from the way V6 treated them.

[MtM]

So you're saying that if I switch my primary screen from my ati card to my nvidia card, my order should change?

I'll try that, interesting if true, or better said worrying ( as what happens when I set arguments for a slot tied to a card with gpu-index in that case ).

Edit:

Didn't change anything with the cards order in the system tab. Which is good, and makes sense.

The cards are listed numerically on vendorID and deviceID, they are not listed in order of active display, enabeld or disabeld cards, primary screen and secondary screen or anything else I can think of. At least, that's why my eyes are telling me.

Run --lspci to check why the 'system' tab list them in the order it does.

[JimF]

MtM,

I expect that if you disabled one driver or another, the order could change, but it looks like you are on Linux, so I don't know. I would have expected that the gpu:x (x=0, 1, etc) designation would stay constant for each card, even if another parameter (e.g., ID number) changed. But I am OK in any case now that I know the convention, as long as the order doesn't change for some reason.

[MtM]

I'm not sure I understand exactly what you mean so let's try again:

Code: Select all

[6/12/2011] - 9:04:37 PM - Searching for configuration file.
[6/12/2011] - 9:04:38 PM - Reading Config.xml file.
[6/12/2011] - 9:04:38 PM - Reading client configuration: True
[6/12/2011] - 9:04:38 PM - Client config report
Remote command server, port: 36331 Password length(!): 6
User configaration, name: Marvin_The_Martian Team: 11314 passkey present(!): True
Number of folding slots: 3
Slot: 0 - SMP
-pause-on-start: true
Slot: 1 - GPU
-core-priority: low
-gpu-index: 2
-pause-on-start: true
Slot: 2 - GPU
-core-priority: low
-extra-core-args: -gpu 0 -forcegpu ati_r700
-gpu-index: 0
-pause-on-start: true
User count: 1
Additional configuration sections: 4
 FahCore Control
-cpu-affinity: true
 Folding Slot Configuration
-gpu: true
 Logging
-verbosity: 4
 Network
-proxy: :8080

[6/12/2011] - 9:04:42 PM - Running -lspci: True
[6/12/2011] - 9:04:42 PM - NVIDIA:
10de:0622:nVidia Corporation:G94 [GeForce 9600 GT]
ATI:
1002:954f:ATI Technologies Inc:RV710 [Radeon HD 4350]
1002:aa38:ATI Technologies Inc:RV710/730

[6/12/2011] - 9:04:45 PM - Client info:
[6/12/2011] - 9:04:45 PM - FAHCLient --info
Website: http://folding.stanford.edu/
Copyright: (c) 2009,2010 Stanford University
Author: Joseph Coffland <joseph@cauldrondevelopment.com>
Args: --info
Version: 7.1.24
Date: Apr 6 2011
Time: 21:37:58
SVN Rev: 2908
Branch: fah/trunk/client
Compiler: Intel(R) C++ MSVC 1500 mode 1110
Options: /TP /nologo /EHa /wd4297 /wd4103 /wd1786 /Ox -arch:SSE2
Platform: win32 Vista
Bits: 32
Mode: Release
OS: Microsoft Windows 7 Ultimate
CPU: Intel(R) Core(TM)2 Quad CPU Q6600 @ 2.40GHz
CPU ID: GenuineIntel Family 6 Model 15 Stepping 11
CPUs: 4
Memory: 8.00GiB
Free Memory: 3.96GiB
Threads: WINDOWS_THREADS
GPUs: 3
GPU 0: ATI:3 RV710 [Radeon HD 4350]
GPU 1: ATI:3 RV710/730
GPU 2: NVIDIA:1 G94 [GeForce 9600 GT]
CUDA: 1.1
CUDA Driver: 4000
On Battery: false
UTC offset: 2
PID: 7524
CWD: C:/Program Files (x86)/FAHClient
Win32 Service: false

[6/12/2011] - 9:04:45 PM - Closing diagnostics


gpu-index corresponds to the GPU 0 GPU 1 and GPU2 in the FAHClient --info section.
-gpu x tells the fahcore which ordinal it should try to initiate ( eg, the card index for the specific runtime ( cal/cuda/opencl platform ).

-gpu x is indeed likely to change with active display, if you set it to -1 it will bind to the first available adapter ( which is always first the primary output ), and if you set it to 0 or 1 and then change which display is the main display, that would indeed swap the card tied to the slot.

gpu-index never changes.

This page isn't correct therefore, gpu-index is not the same as -gpu x ( https://fah-web.stanford.edu/projects/F ... ncesV6ToV7 )

There is another guide/page which explains that gpu-index and -gpu x don't always match and explains it with 'cores', I can't find the page right now but even that is wrong I think. Infact, FAHClient can fold mixed ati/nvidia setups if you override the built in -gpu x increments, which it doesn't do out of the box because it doesn't take into account the different runtimes targeted by each vendor.

[JimF]

-gpu x is indeed likely to change with active display, if you set it to -1 it will bind to the first available adapter ( which is always first the primary output ), and if you set it to 0 or 1 and then change which display is the main display, that would indeed swap the card tied to the slot.

My situation is not that I change the active display; it is always connected to the card in the top PCIe slot, and the bottom card is always unconnected. The confusion is just that the GPU:0 and GPU:1 order is reversed from what it was in V6. Now that I know that, it is not necessarily a big deal, as long as it doesn't go wandering around after that. We will see.

[bruce]

See my theory here: viewtopic.php?f=67&t=18866&p=188828#p188828

[JimF]

Sounds good to me, though on Win7 I don't have lspci as such. But it is said that "Hardware Ids" in the Display adapters properties gives comparable information. At least my two GT 240s are from different manufacturers, so they have different info there, which hopefully will keep them straight.

[MtM]

lspci = 'FAHClient.exe --lspci'

Not nix lspci.

[JimF]

Thank you, this is what I get:

0x10de:0x0ca3:nVidia Corporation:GT215 [GeForce GT 240]
0x10de:0x0ca3:nVidia Corporation:GT215 [GeForce GT 240]

They are identical, unless there is some other level of identification that is missing.

[MtM]

Well I posted it since it should explain the order of the cards, it's more clear if you take my output.

Code: Select all

VendorID:DeviceID:Vendor Name:Description
0x1002:0x954f:ATI Technologies Inc:RV710 [Radeon HD 4350]
0x1002:0xaa38:ATI Technologies Inc:RV710/730
0x10de:0x0622:nVidia Corporation:G94 [GeForce 9600 GT]
0x11ab:0x4320:Marvell Technology Group Ltd.:☺
0x11ab:0x4364:Marvell Technology Group Ltd.:
0x11ab:0x6121:Marvell Technology Group Ltd.:server
0x8086:0x244e:Intel Corporation:
0x8086:0x2e20:Intel Corporation:►
0x8086:0x2e21:Intel Corporation:ä
0x8086:0x2e29:Intel Corporation:
0x8086:0x3a16:Intel Corporation:
0x8086:0x3a20:Intel Corporation:
0x8086:0x3a26:Intel Corporation:
0x8086:0x3a30:Intel Corporation:\
0x8086:0x3a34:Intel Corporation:c
0x8086:0x3a35:Intel Corporation:¿‼f¡Ä¿².
0x8086:0x3a36:Intel Corporation:em
0x8086:0x3a37:Intel Corporation:i
0x8086:0x3a38:Intel Corporation:ddis♥
0x8086:0x3a39:Intel Corporation:
0x8086:0x3a3a:Intel Corporation:
0x8086:0x3a3c:Intel Corporation:
0x8086:0x3a3e:Intel Corporation:
0x8086:0x3a40:Intel Corporation:
0x8086:0x3a48:Intel Corporation:
0x8086:0x3a4a:Intel Corporation:

VendorID goes from low to high, DeviceID goes from low to high ( hexadecimal values ). The fields are actually longer and contain unique numbers, the client should be able to tell them apart but I recon it's not displayed since this output has been added during development to facility easier adding/removing devices from the white list.

[7im]

Yes, V7 does detect GPUs differently. In most cases, it is opposite of v6. Not wrong, just different.