triple 2560x1440 with dual Nvidia GPUs and hardware accel?

Message

maverick256 · Post by **maverick256** » Wed Sep 12, 2012 5:24 pm

Hello, so with all the cheap Korean 27" panels I want to do a triple QHD setup.
Currently I have a Nvidia GTX460 running two of those panels (Achieva Shimian). So I'm thinking of the following 2 options for getting 3 panels running:
1. Get a GTX560, and run a dual GPU setup, or
2. Get a GTX670, 660ti, or an AMD card to run 3 monitors.

I would prefer 1 because I use CUDA to do some simulation work, but the question is, can xrandr do multiple GPUs? I see conflicting information regarding xrandr 1.4 about multiple GPUs, and it's not clear to me if the 304 nvidia driver actually even support 1.4 anyways.
I've seen some threads from 2011 and 2009 that have talked about this subject, but it's not clear if they ever got xrandr with hardware acceleration working or just xinerama.

So my question is, is hardware acceleration viable on 3 screens with 2 Nvidia GPUs, without resorting to separate xscreens? Or do I need to go for option 2?

Thanks a lot.

shazeal · Post by **shazeal** » Wed Sep 12, 2012 6:32 pm

The newer nvidia driver 304.37/43 onward should support it. They changed it so the driver exposes all ports rather than forcing you to select only 2, so instead of twinhead you get as many heads as monitors are connected... in theory. As far as actually getting X to use it I am not sure I only have 2 monitors, but the nvidia-settings software should take care of that for you, you would probably want to install the driver manually rather than using portage though as the portage nvidia-settings is only 302.xx version which may not support it?

I couldnt say on the option 1 vs 2, I would myself go for the nvidia 660 over dual gpu just because of the power savings.

Gusar · Post by **Gusar** » Wed Sep 12, 2012 6:52 pm

The nvidia driver doesn't support xrandr1.4 yet, it's not even sure that it will (they're doing a "feasibility study" currently). Three displays on two GPUs is possible with xinerama, but xinerama is old and outdated and comes with a performance penalty.

So I suggest you get you get a Kepler (6xx) card, which supports up to 4 displays on a single GPU. In theory you could still use the old 460 for CUDA-only.

maverick256 · Post by **maverick256** » Wed Sep 12, 2012 8:26 pm

Wait, so xrandr 1.4 does have multi-GPU support? I can't seem to find consistent information... wikipedia says it doesn't, gentoo wiki says it does, and the official project page seems to be down right now.

Gusar · Post by **Gusar** » Wed Sep 12, 2012 8:42 pm

maverick256 wrote:Wait, so xrandr 1.4 does have multi-GPU support?

Yep: http://cgit.freedesktop.org/xorg/proto/ ... [quote]1.4 Introduction to version 1.4 of the extension

Version 1.4 adds an optional Border property.

• An optional Border property. This property allows a client to
specify that the viewport of the CRTC is smaller than the active
display region described its mode. This is useful, for example,
for compensating for the overscan behavior of certain
televisions.

Version 1.4 adds a new object called a provider object. A provider object
represents a GPU or virtual device providing services to the X server.
Providers have a set of abilities and a set of possible roles.

Provider objects are used to control multi-GPU systems. Provider roles can
be dynamically configured to provide support for:

1) Output slaving: plug in a USB device, but have its output rendered
using the main GPU. On some dual-GPU laptops, the second GPU isn't
connected to the LVDS panel, so we need to use the first GPU as an output
slave for the second GPU.

2) offload - For dual-GPU laptops, allow direct rendered applications to be run
on the second GPU and display on the first GPU.

3) GPU switching - Allow switching between two GPUs as the main screen
renderer.

4) multiple GPU rendering - This replaces Xinerama.[/quote]

maverick256 · Post by **maverick256** » Wed Sep 12, 2012 11:43 pm

hm ok. Thanks so much! I might just wait a bit for now....

hedmo · Post by **hedmo** » Thu Sep 13, 2012 4:09 am

maverick256 wrote:Hello, so with all the cheap Korean 27" panels I want to do a triple QHD setup.
Currently I have a Nvidia GTX460 running two of those panels (Achieva Shimian). So I'm thinking of the following 2 options for getting 3 panels running:
1. Get a GTX560, and run a dual GPU setup, or
2. Get a GTX670, 660ti, or an AMD card to run 3 monitors.

I would prefer 1 because I use CUDA to do some simulation work, but the question is, can xrandr do multiple GPUs? I see conflicting information regarding xrandr 1.4 about multiple GPUs, and it's not clear to me if the 304 nvidia driver actually even support 1.4 anyways.
I've seen some threads from 2011 and 2009 that have talked about this subject, but it's not clear if they ever got xrandr with hardware acceleration working or just xinerama.

So my question is, is hardware acceleration viable on 3 screens with 2 Nvidia GPUs, without resorting to separate xscreens? Or do I need to go for option 2?

Thanks a lot.

Maverick256

I do not know if nvidia does it but I am running 3x LG4200 FLATRON with 2304x1360 and have hardware acceleration on my 2x XFX 6870 in crossfire http://imageshack.us/f/253/photo201208261933471.jpg/

.i dont use xinerama=no hardware acceleration.there is no problem with xrandr,it is setting the screens up just as ati-drivers "fglrx" tells it to and sets them as one screen at 2304x1360.

regards hedmo

maverick256 · Post by **maverick256** » Fri Sep 14, 2012 11:12 pm

whoa, that looks massive! What game is that?
Now I do wonder... doesn't anyone know if dual GPUs with 3 panels can work with Nvidia SLI?

Gusar · Post by **Gusar** » Sun Sep 16, 2012 11:36 am

SLI is very, very poorly supported in Linux. I wouldn't go that way.

If I were you, I'd go with what I already mentioned - buy a Kepler card (the recently released 650 or 660 are good choices) to drive the three displays, and maybe keep the 460 in the machine for CUDA-only.

maverick256 · Post by **maverick256** » Sun Sep 16, 2012 2:57 pm

Yeah you are right, I guess I have to go for the 660, since I can't find any 650 with displayport, and I'm not sure if HDMI can support QHD. I can even evaluate how well the new kepler architecture works with my simulation. Given the specs I would expect it to run a bit slower than on the 460, but should be interesting to see. Thanks!

hedmo · Post by **hedmo** » Mon Sep 17, 2012 5:11 am

maverick256 wrote:whoa, that looks massive! What game is that?

Skyrim.and i have stable 30-40fps overall

rgards hedmo

maverick256 · Post by **maverick256** » Wed Nov 21, 2012 6:12 pm

I don't like to revive an old thread, but just want to show you guys what came out of this whole affair:

http://flic.kr/p/dv6zrn
http://flic.kr/p/dvck3Y

Total cost: about $1330:
3x Achieva Shimian, $900
3x Wennow 27in stand: $150
1x Bizlink active DP to DL-DVI adapter: $50
1x Nvidia GTX 660, $230

On a side note, my CUDA simulation runs about 1.5x faster on the 660 vs the 460.

Obviously that 660 won't be able to handle gaming at 7680x1440...

Nvidia driver's twinview (they might want to rename that) worked pretty easily. The only quirk with these panels is that I had to specify ignore EDID checksum.

Thanks everyone for your advice!