Quad-Monitor AMD/NVIDIA Linux Gaming: What You Need To Know

[Ericg]

AMD GPUs only have two PLLs which means we can only support two independent non-DP clocks. If you want to use more than two monitors, the additional monitors have to be DP. If you try again with 2 DVI/HDMI and 2 DP monitors it should work. Catalyst enforces this requirement. On radeon we allow PLL sharing where it's possible (e.g., multiple panels with the same clock) which is why it works there. That has it's own limitations however. E.g., if you try and change to a configuration that requires more than two PLLs, the modeset will fail.

Alex does the PLL sharing affect the performance? Or is Radeon just not good at multi-monitor at this point? Or is the driver (not sure if its RadeonSI or R600g) for the 6950 just not as close to Catalyst for FPS? Just trying to figure out the numbers gap between Radeon and Catalyst on some of these benchmarks

[agd5f]

Alex does the PLL sharing affect the performance? Or is Radeon just not good at multi-monitor at this point? Or is the driver (not sure if its RadeonSI or R600g) for the 6950 just not as close to Catalyst for FPS? Just trying to figure out the numbers gap between Radeon and Catalyst on some of these benchmarks

PLLs don't affect gfx performance. Having a huge desktop does however. Not only are you rendering to a much larger surface, you have more contention in the memory controller for bandwidth (display controllers and the gfx hardware). You can save bandwidth by doing pageflips rather than blit for buffer swaps, but at the moment, the open source drivers don't support non-vsynced pageflipping due to limitations in the xserver. When you are rendering to huge buffers you also tend to run out of vram so buffers may end up ping-ponging between vram and system memory from frame to frame. Improving the buffer placement heuristics in the open source driver would probably help a lot in vram limited cases.

[darkbasic]

You can save bandwidth by doing pageflips rather than blit for buffer swaps, but at the moment, the open source drivers don't support non-vsynced pageflipping due to limitations in the xserver.

Which ones? Afaik Intel already does Async Display Flipping thanks to DRI3.

[Ericg]

PLLs don't affect gfx performance. Having a huge desktop does however. Not only are you rendering to a much larger surface, you have more contention in the memory controller for bandwidth (display controllers and the gfx hardware). You can save bandwidth by doing pageflips rather than blit for buffer swaps, but at the moment, the open source drivers don't support non-vsynced pageflipping due to limitations in the xserver. When you are rendering to huge buffers you also tend to run out of vram so buffers may end up ping-ponging between vram and system memory from frame to frame. Improving the buffer placement heuristics in the open source driver would probably help a lot in vram limited cases.

2 Questions outta that...

1) Is a the 'non vsynced pageflipping' issue present on Wayland?

2) Will 'ping-ponging' be an issue on HSA systems? AMD CPU with an AMD dedicated card. I haven't checked out the HSA spec so I don't know if you can request to be put on graphics RAM vs system RAM or if you really are just saying "I need a buffer with xyz size" and get given back a pointer and youre blind to whether or not its on DDR3 ( / 4) or GDDR5. Just a thought that came up, figured I'd throw it in haha

[lovenemesis]

AMD GPUs only have two PLLs which means we can only support two independent non-DP clocks. If you want to use more than two monitors, the additional monitors have to be DP. If you try again with 2 DVI/HDMI and 2 DP monitors it should work. Catalyst enforces this requirement. On radeon we allow PLL sharing where it's possible (e.g., multiple panels with the same clock) which is why it works there. That has it's own limitations however. E.g., if you try and change to a configuration that requires more than two PLLs, the modeset will fail.

Here's some more reading backing what agd5f saying.

http://www.anandtech.com/show/7400/t...eat-asus-xfx/3

One side question: Does open source radeon driver support 4K display now?

[agd5f]

Which ones? Afaik Intel already does Async Display Flipping thanks to DRI3.

That requires dri3 and present support. The problem in pre-dri3 land is that the xserver doesn't make the swap interval available to the driver so they can decide whether they want to flip or swap. There have been several attempts to fix it, but nothing has ever made it upstream.

[Ericg]

That requires dri3 and present support. The problem in pre-dri3 land is that the xserver doesn't make the swap interval available to the driver so they can decide whether they want to flip or swap. There have been several attempts to fix it, but nothing has ever made it upstream.

I'm assuming though that DRI3 and Present support is on the roadmap for Radeon?

[agd5f]

I'm assuming though that DRI3 and Present support is on the roadmap for Radeon?

Yes it is.

[Ericg]

Yes it is.

Juuuust checking, thanks for answering questions Alex

[agd5f]

Here's some more reading backing what agd5f saying.

http://www.anandtech.com/show/7400/t...eat-asus-xfx/3

One side question: Does open source radeon driver support 4K display now?

The driver side should be fine assuming you have a new enough kernel to properly parse the 4k modes from the monitor.