4 ms·

I appreciate it's just an industrial design concept, but the problem with off-the-shelf GPUs is that you need to route the DisplayPort connectors back to the mo

I appreciate it's just an industrial design concept, but the problem with off-the-shelf GPUs is that you need to route the DisplayPort connectors back to the motherboard in order to mux them with the Thunderbolt ports. The Thunderbolt add-in cards have DisplayPort inputs for this purpose. The GPU & motherboard vendors should agree on some extra headers to allow you to route these DP signals without ugly jumper cables on the outside of the case.

kogepathic 10y ago
> but the problem with off-the-shelf GPUs is that you need to route the DisplayPort connectors back to the motherboard in order to mux them with the Thunderbolt ports If the Intel CPU supports integrated graphics (which I realize a Xeon may not) there is actually support for sending display streams over PCIe, so it wouldn't be necessary to route DisplayPort cables from the card output to the motherboard. The eGPU guys are doing this to display the output of an external GPU on the internal LCD of the laptop. [0] It does slightly reduce the PCIe bandwidth available as you need to transfer the rendered data back via the PCIe, which is then output from the IGP to the display. [1] However this is still limiting because it means you're stuck with the display outputs supported by the IGP, so if you want to upgrade the video card to support a newer standard (e.g. HDMI 2.0) you'd be stuck. [0] https://www.techinferno.com/index.php?/forums/topic/9658-egpu-experiences-version-20/ https://www.techinferno.com/index.php?/forums/topic/9658-egp... [1] https://www.asus.com/support/faq/1006857 https://www.asus.com/support/faq/1006857 [2] https://www.techinferno.com/index.php?/forums/forum/83-diy-e-gpu-projects/ https://www.techinferno.com/index.php?/forums/forum/83-diy-e...
- xenadu02 10y ago
  Have you tried this? The amount of bandwidth consumed is non-trivial and especially impacts compute workloads or anything that does read-back from the GPU.
  - kogepathic 10y ago
    > Have you tried this? I have not used this feature myself because I'm not aware that it's supported in Linux, and I don't use Windows. > The amount of bandwidth consumed is non-trivial and especially impacts compute workloads or anything that does read-back from the GPU. Given that the most typical use case of this technique is for people gaming, I'd say that the amount of data being read back from the GPU in a gaming situation is quite small, so this doesn't have a significant performance impact for its intended audience. Since I don't use my dGPU for anything but gaming, do you have any more information on the performance impact on compute workloads?
- jamesfmilne 10y ago
  A single 4k/60p/24bit display requires 1.5GB/sec, about 10-15% of your PCIe readback bandwidth. Since it has to be both written to and read back out of main memory, that's 3GB/sec, which is between 5-10% of your CPUs memory bandwidth. It's OK for 1080p res but falls down at higher res. Apple don't do this on their laptops either, and still employ a gmux to switch between integrated and discrete graphics for the laptop display (although that's mainly for power-saving).
cptskippy 10y ago
I'd like to see the industry move away from having connectors on the back of the GPU at all and just provide a standard routing interface of some sort. I hate whenever I buy a new GPU having to figure out what adapters I'll need for my multi-monitor setup because my new GPU has different numbers of HDMI and DP outputs than my old one.
- gambiting 10y ago
  But then you are limited by the number of connectors on your motherboard. I know people who have more cards to support more monitors(not in SLI, just side by side)
- potatolicious 10y ago
  This seems unlikely to happen for the foreseeable future - if ever. As pixel counts grow the port standards are evolving quickly (both HDMI and DisplayPort). A machine where the display interface is on the motherboard will have a hard time keeping up with these developments, whereas keeping the interface on the GPU will allow the user to add support for new standards over time (see: HDR, 8K, etc).

kogepathic 10y ago

> but the problem with off-the-shelf GPUs is that you need to route the DisplayPort connectors back to the motherboard in order to mux them with the Thunderbolt ports If the Intel CPU supports integrated graphics (which I realize a Xeon may not) there is actually support for sending display streams over PCIe, so it wouldn't be necessary to route DisplayPort cables from the card output to the motherboard. The eGPU guys are doing this to display the output of an external GPU on the internal LCD of the laptop. [0] It does slightly reduce the PCIe bandwidth available as you need to transfer the rendered data back via the PCIe, which is then output from the IGP to the display. [1] However this is still limiting because it means you're stuck with the display outputs supported by the IGP, so if you want to upgrade the video card to support a newer standard (e.g. HDMI 2.0) you'd be stuck. [0] https://www.techinferno.com/index.php?/forums/topic/9658-egpu-experiences-version-20/ https://www.techinferno.com/index.php?/forums/topic/9658-egp... [1] https://www.asus.com/support/faq/1006857 https://www.asus.com/support/faq/1006857 [2] https://www.techinferno.com/index.php?/forums/forum/83-diy-e-gpu-projects/ https://www.techinferno.com/index.php?/forums/forum/83-diy-e...

xenadu02 10y ago
Have you tried this? The amount of bandwidth consumed is non-trivial and especially impacts compute workloads or anything that does read-back from the GPU.
- kogepathic 10y ago
  > Have you tried this? I have not used this feature myself because I'm not aware that it's supported in Linux, and I don't use Windows. > The amount of bandwidth consumed is non-trivial and especially impacts compute workloads or anything that does read-back from the GPU. Given that the most typical use case of this technique is for people gaming, I'd say that the amount of data being read back from the GPU in a gaming situation is quite small, so this doesn't have a significant performance impact for its intended audience. Since I don't use my dGPU for anything but gaming, do you have any more information on the performance impact on compute workloads?
jamesfmilne 10y ago
A single 4k/60p/24bit display requires 1.5GB/sec, about 10-15% of your PCIe readback bandwidth. Since it has to be both written to and read back out of main memory, that's 3GB/sec, which is between 5-10% of your CPUs memory bandwidth. It's OK for 1080p res but falls down at higher res. Apple don't do this on their laptops either, and still employ a gmux to switch between integrated and discrete graphics for the laptop display (although that's mainly for power-saving).

cptskippy 10y ago

I'd like to see the industry move away from having connectors on the back of the GPU at all and just provide a standard routing interface of some sort. I hate whenever I buy a new GPU having to figure out what adapters I'll need for my multi-monitor setup because my new GPU has different numbers of HDMI and DP outputs than my old one.

gambiting 10y ago
But then you are limited by the number of connectors on your motherboard. I know people who have more cards to support more monitors(not in SLI, just side by side)
potatolicious 10y ago
This seems unlikely to happen for the foreseeable future - if ever. As pixel counts grow the port standards are evolving quickly (both HDMI and DisplayPort). A machine where the display interface is on the motherboard will have a hard time keeping up with these developments, whereas keeping the interface on the GPU will allow the user to add support for new standards over time (see: HDR, 8K, etc).