GPU Low Profile GPU Potential/Discussion

Analogue Blacksheep

Master of Cramming
Original poster
Bronze Supporter
Dec 2, 2018
483
331
I've been recently looking at low profile GPUs; the 1050ti quadruplets, the P1000 and the WX4000, and it got me thinking: Is there any potential for low profile GPUs to get as powerful as their ITX counterparts with the current technology? Say RTX 2070 performance levels?

I was wondering if anyone here knew the reasons as why low profile GPU's seem to be overlooked/underdeveloped. Technical limits or a lack of interest in the market?
 
Last edited:

cleveland

Master of Cramming
Sep 8, 2016
431
226
I was wondering if anyone here knew the reasons as why low profile GPU's seem to be overlooked/underdeveloped. Technical limits or a lack of interest in the market?
It may sound like a meme, but "why not both"?

Also, have you looked at the "guts" of a GTX 1070? Just for comparison purposes, the GT 1030 have almost 2 billion transistors count. The simplest 1070 model have as much as 7.2 billion transistors. More than 3 times the transistor density, in a relative comparison!

Now think about the thermals. Physics can't be set aside, despite consumer's wishes. 7 times the power draw means 7 times the heat output (roughly), in less than half the dissipation area. That's a lot of heat, even for SFF enthusiasts.

IF (and that's a pretty big "if") we can reach 1060's performance within half-height specs, I'd ve VERY surprised, even with aditional power connectors and optimized dissipation technology.
 

NateDawg72

Master of Cramming
Aug 11, 2016
396
283
We should see a new generation of LP cards when 7nm gets used in more GPUs. They'll be a good step up from current LP cards, but it's unlikely that the core limitations will change. Size and power are the technical problems that prevent larger chips like RTX 2070 being used on a LP card. Larger "ITX" sized cards will still end up faster from having the room to use a larger GPU and all the necessary VRAM and power delivery hardware.
 

Valantar

SFF Guru
Jan 20, 2018
1,353
1,144
LP GPUs are (generally) a similar length to ITX full-height GPUs (as this matches the HHHL standard size), with dual slot coolers. ITX GPUs are quite often taller than official PCIe AIC height to maximize fan size and fin stack size. As such, at the very best, a low-profile card can cool <1/2 the wattage of a full height ITX card. Then there are power delivery issues: the slot only delivers a maximum of 75W (though this is actualy a bit of an overstatement). Adding a 6-pin PCIe connector would eat up a significant portion of board space on such a small card, which would reduce cooling potential. This jives pretty well with current available offerings: the most powerful ITX cards approach 180+W, while the most powerful ITX cards reach about 75W - which is both the power limit and where cooling potential seems to peak (-ish).

Making a three-slot LP card could alleviate both of these factors (but also introduce a host of compatibility issues), but then you get in trouble in terms of actually fitting hardware on the card - the package of most high-end GPUs would pretty much span the entire height of an LP card, and then you have to fit RAM (needs a lot of space!) and VRMs (also need quite a bit of space). A card with a reasonably sized die and HBM would likely fit, but power delivery would still be a challenge, as would routing heat pipes for such a heavy heat load within the tight confines of a HHHL area. Some sort of dual-board sandwich layout (like some dual-GPU cards way back when) would alleviate this, but would have a serious impact on cooling potential (smaller heatsinks, only radial fans) and would dramatically increase costs.

In other words: it could be possible, but it would be exorbitantly expensive, incompatible with the vast majority of low-profile cases, and have absolutely terrible value for money.
 

robbee

Master of Cramming
Silver Supporter
Sep 24, 2016
468
572
Adding a 6-pin PCIe connector would eat up a significant portion of board space on such a small card, which would reduce cooling potential.
This has been done often in de past though: I remember seeing the GTS 450 and 9600 GT in low profile with 6pin PCI connectors. Also low profile cards longer than 170mm seemed to be pretty normal back then. This form factor would certainly allow for interesting builds.
 

Valantar

SFF Guru
Jan 20, 2018
1,353
1,144
Would PCIE 4 help low profile GPU's in anyway? Despite the fact PCIE 4 is keeping the 75W power limit. There was a rumour back in 2016 suggesting it was going to go up 300W.
A power increase like that would make motherboards significantly more expensive (as they'd likely need additional copper layers to handle that much power, as well as an 8-pin PCIe power connector per x16 slot) which might be why this didn't pan out (if the rumor had any grain of truth to it at all). Other than that PCIe 4.0 shouldn't have any effect - even 2080tis aren't limited by PCIe x16, let alone anything bus-powered or that can fit in a HHHL package.
 

NateDawg72

Master of Cramming
Aug 11, 2016
396
283
The spec did increase to 300W (up from 250W I believe) but the slot is still 75W - the rumor was a misinterpretation. Additional power connectors are needed to reach the 300W specification, as it is a total board power specification.
 
  • Like
Reactions: Valantar

Choidebu

"Banned"
Aug 16, 2017
1,133
1,120
So if I'm understanding this correctly, LP cards are limited by their heat dissipation real estate?

If, a big IF, a manufacturer decided to make a water cooled only LP card, preinstalled waterblock with common fittings for aftermarket loop, it is possible?
 

Necere

Shrink Way Wielder
NCASE
Feb 22, 2015
1,652
3,137
This discussion reminds me of a concept for a triple-slot low-profile card I once did:



You can tell how old this is by the GPU series. Those are 50mm fans up front, so they would be really more suited to a server or something where you didn't care about noise.

Expanding on this idea a bit (literally), moving to a quad-slot design could allow for an upward-facing blower fan:



Compared to a reference blower, this would be a bit smaller in diameter, but could be significantly thicker.


Alternatively, another idea I had was to expand to the cooler to two slots, but use the extra space below the card for a larger cooler:



Either a blower or fans up to ~80mm on an open cooler could work here. This would limit the card's compatibility to mini-ITX only though, as the cooler "spillover" would conflict with larger boards.
 
  • Like
Reactions: Analogue Blacksheep

Valantar

SFF Guru
Jan 20, 2018
1,353
1,144
So if I'm understanding this correctly, LP cards are limited by their heat dissipation real estate?

If, a big IF, a manufacturer decided to make a water cooled only LP card, preinstalled waterblock with common fittings for aftermarket loop, it is possible?
Partially - you'd still have the challenge of fitting a large GPU package, the necessary VRAM, and the VRMs onto a very small card. For example, the TU106 would be extremely difficult to fit due to its size. Non-RTX TU-series chips would likely be easier to fit, but the necessary VRAM for a modern mid-range GPU (and the bandwidth needed) makes even that very difficult.
 
  • Like
Reactions: NateDawg72

Valantar

SFF Guru
Jan 20, 2018
1,353
1,144
I made a couple of photoshop mockups to illustrate the problem, with an LP 1050Ti combined with a 1660Ti (the smallest Turing die/package).



So, what we can immediately see is that a) the package and die are dramatically larger than the 1050, and while it could technically fit on the board, attaching its required six VRAM channels would be impossible - there simply isn't room for the chips, let alone the traces. Then there's the VRM requirements, and the power connector.

Long story short, you need a smaller die and package for LP GPUs.
 

Valantar

SFF Guru
Jan 20, 2018
1,353
1,144
Or just thicker plates and better engineering skills
Thicker plates? You're going to have to explain what that means to me. You mean a thicker PCB? Sure, that could shrink some things, but likely not the GPU package substrate, which is the biggest component. It's sized for pinout/BGA pitch requirements and shrinking that would require a more complicated and expensive production process. RAM packages still take the same space, and fitting more than 4 seems near impossible. More PCB layers could allow for denser DRAM traces, but trace length matching (needed for timings to match and thus stability) is still a requirement. Multi-layer DRAM traces also come with added risk of interference. A thicker PCB might allow for adding a couple of DRAM packages to the back of the card, though, which would make this a bit easier. It would still be disproportionately expensive.
 

Thehack

Spatial Philosopher
Creator
Bronze Supporter
Mar 6, 2016
2,659
3,352
J-hackcompany.com
yes

All the rest you wrote is tied to better engineering skills. Thanks to validating my point.
It is not due to lack of engineering though. There are physical constraints when it comes to very precise electrical signals. There are physical limitations to how you can lay a board out.

It would be possible if it used HBM but that's in another world. We got what we got.
 

Valantar

SFF Guru
Jan 20, 2018
1,353
1,144
yes

All the rest you wrote is tied to better engineering skills. Thanks to validating my point.
Neither cost nor unavoidable physical constraints are synonymous with "engineering skills". Even disregarding cost (which no OEM will do for a low-margin, mid-price product like a ~$220-250 GPU), it's highly doubtful that it would be possible to route DRAM to the back of the PCB (due to other SMDs needed behind the package, location of DRAM pins on the package (and thus their routing - DRAM packages are usually placed along the side of the die where their traces originate), and signal interference issues with crossing/parallel DRAM traces on different PCB layers. And you'd still need to fit the 1660's increased VRM needs over the 1050 into the same area - again not impossible, but either hot or expensive (or both).
 

cleveland

Master of Cramming
Sep 8, 2016
431
226
Neither cost nor unavoidable physical constraints are synonymous with "engineering skills".
Cost is a matter of markets. If there is demand, there will be supply, without even the slight doubt.

"Phisycal constraints" is what didn't impair engineers from designing 7nm GPU architectures, starting from the ancient GeForce 256.

One thing i can say for sure: we won't have a newer half height GPU with that attitude.