AMD's Initial Radeon Driver Changes For Linux 3.12

[agd5f]

Ah ok, I thought multi-gpu was also under the umbrella of eyefinity. Is there a marketing name for this?

not sure. hydravision maybe?

[CalcProgrammer1]

On Windows (and probably fglrx, but I've never had success with it) Eyefinity only works with monitors from one GPU. You plug all your monitors into one GPU and then the other GPU has no monitors attached. Both GPU's have to have the framebuffer though (maybe GPU with monitors has the whole framebuffer while secondary GPU only has the part it's rendering, not sure). If they could make it so SLS surfaces could be made that span multiple GPU's that'd be awesome for video walls and such. When they first demoed Eyefinity they had like a 24 monitor video wall, never was explained how that worked since 24 monitors across 4 GPU's wasn't an option in Windows drivers.

[agd5f]

There is one bug right now: You need to change the driver name in the ~/.drirc file it makes from the name of the driver (r600, etc) to dri2 for it to work.

This was fixed a while ago. Only older versions of mesa require dri2 for the driver name for gallium drivers.

[Kivada]

You don't need Crossfire for OpenCL because the API explicitly supports multiple GPUs. You need Crossfire for graphics because most usage scenarios need multiple GPUs to appear as a single GPU to the API.

Can it be limited to a single GPU? I was thinking my next system will be whatever the top end Kaveri APU will be with a Hawaii serie dedicated GPU. I want the GPU in the APU to run just as an OpenCL processor.

Reason? Video transcoding while gaming without skipping a beat.

Also do you have any info on the AMD Seattle ARM server SoCs? Like what kind of GPU performance will they have and will they have OSS drivers? Will they be packaged in a way that they can be used in something like a laptop out of them? Low power draw HTPC front end box? The PCIe they are stated to have, will it support off the shelf PCIe cards?

**These are the kinds of questions Larabel should be asking...

[eduperez]

After many "NVIDIA-powered" boxes, something tells me my next computer will have an ATI card inside...

[bridgman]

The fglrx feature to spread a single surface across multiple GPUs was called Multiview. It's pretty complex, primarily because every drawing operation has to be checked to see which GPU(s) it needs to be rendered on. IIRC the equivalent capability for open source 2D (EXA) was called "shatter", and I don't remember much discussion about 3D (OpenGL).

In general it ends up faster doing all the rendering on one GPU into a single large frame buffer and copying portions to other GPUs for display as Alex said -- the time you lose on the copy you make up by not having to check and possibly replicate all the drawing operations.

[droste]

When they first demoed Eyefinity they had like a 24 monitor video wall, never was explained how that worked since 24 monitors across 4 GPU's wasn't an option in Windows drivers.

That's why I thought multi-gpu is called Eyefinity. I've done that with 2 GPUs and 10 monitors in windows (with 1 application spanning all 10 monitors). And Later this month or next month I have to test if 4 GPUs and 24 monitors work

[bridgman]

That's why I thought multi-gpu is called Eyefinity. I've done that with 2 GPUs and 10 monitors in windows (with 1 application spanning all 10 monitors). And Later this month or next month I have to test if 4 GPUs and 24 monitors work

That demo was 4 Eyefinity setups each with 6 monitors displaying from a single large surface, and the app rendered to 4 surfaces.

[Kivada]

...

So I take it it's still too far out for you to make any statements on the ARM based chips?

[bridgman]

So I take it it's still too far out for you to make any statements on the ARM based chips?

Yeah, we do not comment on unreleased products blah blah blah. More specifically, the marketing folks decide what info should be released and we stay inside those lines. I can say that AMD has always been pretty aggressive about open source Linux support for server parts and there are no plans to change that.