nMP Display Ports and GPU Mappings

[VirtualRain]

Interestingly, I've been playing around with my two 24" LED Apple Displays on my nMP and it doesn't seem to matter what TB ports I use to connect them, System Report always shows both of them connected to the GPU in Slot 2. See attachment.

I wonder if this is the nMP's way of keeping one GPU dedicated for compute while the other one is dedicated to driving displays.

Related to this... In this Apple KB article, it says...

Note: Attach displays to different Thunderbolt busses when possible (see figure below). Do not attach more than two displays to any Bus. This means that if you use the HDMI port, be sure to then only use one of the bottom two Thunderbolt ports (Bus 0).

Here are the TB bus configurations... (taken from the article)

I assume this is really for 4K display users since there's not enough BW on a single TB bus to drive dual 4K displays. I've got both my MDP displays connected to one bus at the moment and it's not a problem (although my displays only consume about 6Gbps at 1920x1200x32@60Hz).

Anyone have any other theories or comments on the way Apple has implemented display routing to GPUs in this machine?

 

[Einz]

I noticed that too. Now i don't have to waste my time looking for slot 1.

Thanks.

 

[jav6454]

Bandwidth is key here. 4k resolutions are quite heavy in terms of bandwidth consumption.

Three main channels that divide into three.

 

[Cubemmal]

It makes sense to separate out displays. Except this comment

Do not attach more than two displays to any Bus.

That's pretty definitive. I presently have two TB Displays connected to my 2011 17" MBP. Why can't (?) I connect those two to a single port of the nMP?

This means that if you use the HDMI port, be sure to then only use one of the bottom two Thunderbolt ports (Bus 0).

More confusion, what do they mean here? If they mean "be sure to only use one (Display) of the bottom two Thunderbolt ports", then we have two displays connected to one bus. If they mean ANY TB device, then why only use one port? I have no idea what they're trying to say here.

Anyhow, Anand couldn't figure out the video muxing either, and he's the one person that could figure it out. They have a PCIe lane switch, presumably they have a TB video switch too.

Why don't you post the extended System Report of the TB chain on this thread so we can poke through to see if there's any kind of switch?

 

[VirtualRain]

It makes sense to separate out displays. Except this comment



That's pretty definitive. I presently have two TB Displays connected to my 2011 17" MBP. Why can't (?) I connect those two to a single port of the nMP?




More confusion, what do they mean here? If they mean "be sure to only use one (Display) of the bottom two Thunderbolt ports", then we have two displays connected to one bus. If they mean ANY TB device, then why only use one port? I have no idea what they're trying to say here.

Anyhow, Anand couldn't figure out the video muxing either, and he's the one person that could figure it out. They have a PCIe lane switch, presumably they have a TB video switch too.

Why don't you post the extended System Report of the TB chain on this thread so we can poke through to see if there's any kind of switch?

Displays show up under Graphics/Displays. The Thunderbolt Bus shows only data devices... so you can't tell anything about how displays are routed from that. FWIW, here's the situation with my 3 TB buses... (as you can see I only have my WD TB Duo attached that's a data peripheral)...

 

[analog guy]

virtualrain needs to update his sig!

 

[VirtualRain]

virtualrain needs to update his sig!

Done!

 

[VirtualRain]

The more I think about this and read up on AMD display support, the more I believe that only a single GPU is used to drive any and all displays in the nMP... either six regular displays or three 4K displays (and combinations in between). That means the other GPU in the nMP is not doing any display duty and is strictly for compute.

Evidence that would support this...

- Anand stated this...

By default, one GPU is setup for display duties while the other is used exclusively for GPU compute workloads.

- A single 7xxx series GPU is spec'd for...

Up to 6 displays supported with DisplayPort 1.2 Multi-Stream Transport

- No matter which TB ports I connect my displays to, they end up connected to the same GPU

- And, it makes sense that if a GPU can drive 6 regular displays that it can drive three 4K displays with MST (combining two display streams together to drive a 4K display).

My next question is why they would do it this way. Why not split display duty across the two GPUs? does this impact the performance of a GPU on compute tasks (I guess at the very least displays consume VRAM for frame buffer duty)?

EDIT: If this is the case, the importance of VRAM increases dramatically if you intend to do 3D work on 4K displays. Most early tests of 4K 3D gaming show a 2GB frame buffer to be limiting and 3-4GB optimal which could be very indicative of other 3D workloads at 4K resolutions.

 

[deconstruct60]

....
That's pretty definitive. I presently have two TB Displays connected to my 2011 17" MBP. Why can't (?) I connect those two to a single port of the nMP?

You can. A single port of the nMP is not a single "TB bus". Furthermore, it says not hook up 'more than two'. Hooking up just two TB displays is not more than two monitors.

What it says do not do would be two TB displays on either one or both of the bottom two TB ports and hook up a 3rd display on the HDMI port. That would make three displays on TB bus 0.

That is the easy way to three. You can also do three on ports 1-3 or 2-4 by using DisplayPort v1.2 and displays that daisy chain. Or by trying to daisy chain 3 (or more ) TB monitors. Or any combo of TB or DisplayPort daisy chaining. Just because the other "TB bus" just have two ports doesn't mean can't attempt to hook up three monitors to those two. There isn't necessarily a 1:1 ratio of ports to monitors.

 

[deconstruct60]

....

My next question is why they would do it this way. Why not split display duty across the two GPUs?

A. FCPX doesn't deal well with app windows on different GPU cards. If only hook up one card and not fix that issue with the app ... issue firmly swept under the rug?

B. 300 pins on the GPU connectors but if not using all of them (interference/separation) then x4 PCI-e v2 off one card and 6 Display channels off the other? ( i.e., a subset of paths off card that can be used for one or the other. If so that is bad news long term for a 2nd internal SSD option. ]

C. If buy into Crossfire mindset then there are defacto "master" and "slave" cards. Actual display(s) are only hooked to the "master".

does this impact the performance of a GPU on compute tasks (I guess at the very least displays consume VRAM for frame buffer duty)?

I would allow brain dead simple load balancing. And not as much performance but down clocking when can't fill up the compute GPU with workload.

GPGPUs don't multitask extremely well. It isn't as bad as previous generations but not quite "general purpose" either.

 

[VirtualRain]

A. FCPX doesn't deal well with app windows on different GPU cards. If only hook up one card and not fix that issue with the app ... issue firmly swept under the rug?

B. 300 pins on the GPU connectors but if not using all of them (interference/separation) then x4 PCI-e v2 off one card and 6 Display channels off the other? ( i.e., a subset of paths off card that can be used for one or the other. If so that is bad news long term for a 2nd internal SSD option. ]

C. If buy into Crossfire mindset then there are defacto "master" and "slave" cards. Actual display(s) are only hooked to the "master".

All of this makes a lot of sense. I guess that's it... One GPU for displays, one for compute.

 

[Cubemmal]

You can. A single port of the nMP is not a single "TB bus". ...

Ah, I just read it too quickly. Makes sense, two monitors max, but ideally one per bus. And no cheating by using HDMI either.