PCI-e x32?

[xnamkcor]

Is there some kind of addressing problem with having 32 lanes on one slot? I don't think it's power, you can always add auxiliary power connectors.

 

Is it a physical space issue on the PCB? If someone did make a board that supported it and a card that uses it, could we see "x4 GPU"s in one PCB?

[Glenwing]

Well, the slots would be like 60-70% longer. And there's no need for it other than compute, really. It doesn't exist because there is zero demand for it.

[Vitalius]

Is there some kind of addressing problem with having 32 lanes on one slot? I don't think it's power, you can always add auxiliary power connectors.

 

Is it a physical space issue on the PCB? If someone did make a board that supported it and a card that uses it, could we see "x4 GPU"s in one PCB?

x4 GPU already exist in commercial environments. That's what an Nvidia Quadro is IIRC. I mean, obviously not for gaming, but still.

[Kyuubixchidori]

be interesting down the road when GPUS need that much bandwidth. maybe they will be using a new connector by then? 

[xnamkcor]

 

Well, the slots would be like 60-70% longer. And there's no need for it other than compute, really. It doesn't exist because there is zero demand for it.

(t-p)*2 longer. Where t is the total length, and p is the length for the power.

If there were "no demand" for x32 slots, then SLI or XF wouldn't exist. And last I checked, We still haven't completely solved micro-stuttering.

 

x4 GPU already exist in commercial environments. That's what an Nvidia Quadro is IIRC. I mean, obviously not for gaming, but still.

But Quadro isn't built for gaming. And that's an important detail for a customer who plays games.

 

 

The main issue I see is either PCB space not being there and having to re-arrange the layout to accommodate for not only space, but EMI too. The other issue is that a single x32 slot would only allow a single card to be used, wasting 16 lanes. Maybe if there could be some way to link two x16 slots, but still let them operate independently, if needed.

[Hunter7263]

x4 GPU already exist in commercial environments. That's what an Nvidia Quadro is IIRC. I mean, obviously not for gaming, but still.

Quadros are just Nvidia cards that have 10 bit color unlocked and are thoroughly tested to work for certain commercial softwares. They also are the the best of the binning process and have the nicest components because they are designed to be run 24/7.

[Glenwing]

 

 

(t-p)*2 longer. Where t is the total length, and p is the length for the power.

If there were "no demand" for x32 slots, then SLI or XF wouldn't exist. And last I checked, We still haven't completely solved micro-stuttering.

 

But Quadro isn't built for gaming. And that's an important detail for a customer who plays games.

 

 

The main issue I see is either PCB space not being there and having to re-arrange the layout to accommodate for not only space, but EMI too. The other issue is that a single x32 slot would only allow a single card to be used, wasting 16 lanes. Maybe if there could be some way to link two x16 slots, but still let them operate independently, if needed.

 

 

Don't know what you mean. SLI and CrossFire exist because people want more GPU power, not because they want PCIe x32 slots.

[Hunter7263]

 

 

(t-p)*2 longer. Where t is the total length, and p is the length for the power.

If there were "no demand" for x32 slots, then SLI or XF wouldn't exist. And last I checked, We still haven't completely solved micro-stuttering.

 

But Quadro isn't built for gaming. And that's an important detail for a customer who plays games.

 

 

The main issue I see is either PCB space not being there and having to re-arrange the layout to accommodate for not only space, but EMI too. The other issue is that a single x32 slot would only allow a single card to be used, wasting 16 lanes. Maybe if there could be some way to link two x16 slots, but still let them operate independently, if needed.

 

Currently it would waste lanes on most set ups. Most processors right now can only do 16 lanes at a time unless you step up to the really nice i7's which can do more. The only possible market right now is commercial stuff (I think some xeons can do more than 16 lanes) and the 2011 socket crowd which is like the top 1% if that. It would be nice in theory but probably not happening anytime soon.

[Vitalius]

@xnamkcor Also note that AMD runs Crossfire through PCI-e in the 290, 390, & probably through the Fury X GPUs. 

[xnamkcor]

@xnamkcor Also note that AMD runs Crossfire through PCI-e in the 290, 390, & probably through the Fury X GPUs. 

In the case of doing it "through PCI-e", how does that differ from using different slots? If you had 2 x8 cards that did Crossfire "through PCI-e", is it any different than having both GPUs on the same PCB and using a 16x slot?

Could you put 2 or 4 GPUs on one PCB and give it 2 x16 interfaces?

[Vitalius]

In the case of doing it "through PCI-e", how does that differ from using different slots? If you had 2 x8 cards that did Crossfire "through PCI-e", is it any different than having both GPUs on the same PCB and using a 16x slot?

Could you put 2 or 4 GPUs on one PCB and give it 2 x16 interfaces?

The main thing I think of is heat. You have them separated so it's easier to cool them individually. There's slightly more latency due to them having to go through the CPU via PCI-E lanes than being right beside each other. 

I'd imagine you could if it were better supported.

[xnamkcor]

The main thing I think of is heat. You have them separated so it's easier to cool them individually. There's slightly more latency due to them having to go through the CPU via PCI-E lanes than being right beside each other.

I'd imagine you could if it were better supported.

What if two PCI-e x16 were fully operational as separate x16 slots, but were also linked do that they didn't have to "go through the CPU"? Even if you had to match cards like you do with dual channel RAM, it might be a solution that would solve most issues.