Modified PCIe 8-pin connector

[jØrd]

desktop mxm probably maybe almost became a thing
http://pclab.pl/art21942-2.html
http://www.bit-tech.net/news/hardware/2006/01/07/geminium_go_mxm_sli/1

 

[GuilleAcoustic]

Liantec has ITX motherboard sporting an ultra compact PCIe-16x, called Tiny-Bus 16x. They use it to provide mezzanine boards, like this MXM3.0 one.

They also have ITX boards with MXM connectors on them.

Sadly, it is limited to core2 era of CPU. Here are the available Tiny-Bus expansion boards: http://www.liantec.com/product/TBM.htm

 

[iFreilicht]

I saw that sometime in the past and it's pretty darn cool, but I feel the uses are extremely limited, even when transferred into the current era of CPUs. Firstly, a socketed CPU including cooling has no space in such a design, so you'd need one soldered to the board, with the cooler already installed. Secondly, you're still strongly limited in terms of power the GPU can use. I guess it would be quite interesting for builds that would normally include a 750Ti, but before something like that is standardised, APUs will have already gained enough performance to make this an expensive and non-viable experiment. Not to say that I wouldn't like to see this happening, you could certainly figure out solutions for all of this, including the cooling, maybe an integrated AIO loop that cools both CPU and GPU could work, but it seems like a very complicated undertaking.

 

[GuilleAcoustic]

Another thing .... A motherboard with two processor sockets: one CPU processor socket and one GPU processor socket, with shared HBM between the two. But that will never happen

 

[iFreilicht]

That is an awesome idea, but it would have huge architectural implications as you're operating a multi-processor system at that point, so it needs to be determined how much the two should be coupled, how the shared memory is handled in terms of updating and so on.
The trick with HBM is that it's on the same package as the processor, so shared HBM can't be slotted, it would have to exist on the CPU and GPU package.

As an intermediate step, why not have two sockets, but the GPU socket fits the whole package with HBM on it, and that socket is connected to the CPU via PCIe x16 4.0? That way, the architectural stuff is sorted out, we "only" need to define a socket and convince everyone to migrate from physical PCIe x16 to that. Easy!

Sorry, this discussion is extremely Off-Topic, I'll stop now.

 

[Phuncz]

The reason FirePro and Tesla/Quadro cards have the PCIe connector on the side instead of on top like most commercial GPUs, is because these cards go in workstations and servers.
Otherwise you'd have this:

4U height

But enterprise wants this:

3U height

 

[|||]

Following the same idea, we need a Sata data + power combo connector on the motherboard side. That shouldn't be too hard to do and a single cable for both would really reduce the number of cables and help with their management while improving hairflow.

Combining SATA data and power seems much more sensible to me, as the number of drives is already limited by the motherboard and they don't need a lot of power anyway. The question is, would such an interface support 5.25" drives as well? Those need more power which would potentially increase the size of the connector, so it might be worth to think about it as an interface only for 2.5" drives and Slimline drives.
That way, you could potentially use something like the eSATAp connector which is quite small and supplies 5V and SATA data, which is already sufficient for slimline drives and should be enough for most SSDs. Maybe it could also be designed in a way that makes 5mm thick 2.5" drives viable again.

SATA is a dying standard anyhow, so I wouldn't put much thought into it. M.2 have power included in the connector as do PCI-e cards. The NVMe PCI-e controllers are still too power hungry to consider put power through the Mini-SAS connectors...perhaps down the road after the power has be reduced in them.

 

[GuilleAcoustic]

Mechanical drives aren't dead yet and I do not see the Sata standard dying before several years.

 

[EdZ]

The NVMe PCI-e controllers are still too power hungry to consider put power through the Mini-SAS connectors...perhaps down the road after the power has be reduced in them.

It;s not the power requirements, SFF-8643 simply lacks the pins required for power delivery (all its pins are used up by signal pairs and their grounding pins). SFF-8639 (AKA u.2) has no problems handling power delivery, it just needs the outboard power connector to actually get power to the interface.

 

[|||]

Right, and neither does SATA. I made that comment in response to the data + combo connector off the motherboard that they were proposing for SATA, just for emerging interconnects that will have lifespans longer than a couple years and could justify the cost and effort to develop.

 

[iFreilicht]

The thing is that we already have an interconnect with SATA that supports that, which is called eSATAp, where the pins of an eSATA and a USB connector are combined in one connector to deliver SATA data and 5V power or just USB2.0. eSATAp didn't get standardised, though.

It would allow a single cable to deliver data and power to all SATA devices that only draw 5V power, and on the device-end, a standardised connector does exist in the form of Slimline SATA. Sure, only slim ODDs use it at this point in time, but 3.3V delivery was already removed from the SATA spec and as 2.5" drives have to work in laptops, which usually can't supply 12V anyway, they probably only use 5V already.

 

[Phuncz]

I believe enterprise HDD drives and stuff like the Velociraptor do use the 12V, due to the 10K+ rpm motor I'd reckon. Not that this matters for the use cases here but just to add some info.

 

[iFreilicht]

That's actually true, there are a lot of forum posts by people that wanted to install the Velociraptor into their Xbox 360 Slim, which only supplies 5V to the SATA connector. All the other enterprise drives are SAS anyway, so those're'nt coun'in', I tell ya.

 

[XeaLouS]

Just wanted to say that these are already being made:
http://www.ebay.com/itm/PCI-E-8-pin...-with-low-profile-4-Made-In-USA-/251029888932

 

[|||]

I've seen those...they're still too high for my application and you can get them even lower profile than what they are offering.

 

[Dyson Poindexter]

I've always been miffed at how the PCIe power connector increases the width or depth of a card in a case. For ultra-custom builds I'd like to see a connector mounted such that it protrudes vertically from the back side of the GPU.

I'm one of the hardcore types that wants a cube case that is no wider/deeper than the motherboard and no taller than the video card.

Something like this "chopped osmi" below.

 

[Phuncz]

You could solder it, most of the time they are just through-holes I think. You could find a straight board-mount PCIe connector online most likely.

 

[iFreilicht]

most of the time they are just through-holes I think.

Always. Molex Mini-Fit doesn't exist as SMD, only as through hole. So yeah, one could try to desolder the connectors that are on the GPU and just solder in straight ones themselves, there's no reason why that wouldn't work.

 

[BirdofPrey]

Do be aware, though, if you solder it to the backside, the connections will be backwards.
It's best to avoid connecting 12VDC to the common ground lines

 

[iFreilicht]

Not necessarily. For 6 pin connectors, it doesn't matter at all as long as you solder the connector the right way around.
With 8pin connectors, you can either solder it so that top and bottom pins are swapped or that it is mirrored right to left, that's true.