GPU Low Profile GPU Potential/Discussion

Valantar

Shrink Ray Wielder
Jan 20, 2018
2,201
2,225
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.
Well, you're just making it clear that were debating with fundamentally different starting points. I was talking about fitting a GTX 1660 onto a HHHL card. The TU116 is a known entity with a given set of features which can't be changed without making a different product entirely. Your line of argumentation seems to be founded on a premise of "what if everything just got better", which it no doubt will, but not in a time frame where the TU116 is still relevant, nor in a way that won't necessitate a new product anyhow. We can all dream of HBM-equipped HHHL GPUs, and a far larger die than TU116 could fit in that form factor with HBM and an interposer - but then we're no longer debating the same topic at all. The possibility of better/bigger chips in HHHL form factors with the use of currently exotic/expensive manufacturing or packaging tech doesn't change the fact that TU116 can't be fit into that form factor without significant compromises (such as a reduction in memory bandwidth).
 

cleveland

Master of Cramming
Sep 8, 2016
455
240
Well, you're just making it clear that were debating with fundamentally different starting points
I don't think so.

"what if everything just got better", which it no doubt will, but not in a time frame where the TU116 is still relevant
your doubt is based on marketing statistics that based on... well, your assumptions about the market as a whole. a few dots ahead, you point that the enginering improvement is simple, easy, well known and - yes - expensive. Comparatively, it is as expensive as purchasing, from the other side of the globe, a single unit of a very niche non consumer grade motherboard... and we know is not absurd, since we do this all the time, here in this same forum.

As a matter of fact, the very existence of HHHL GPUs is a proof that the market dont care so much about price, since these GPUs sit a bit behind its non-HHHL counterparts, considering raw performance and thermals.
 

Valantar

Shrink Ray Wielder
Jan 20, 2018
2,201
2,225
I don't think so.


your doubt is based on marketing statistics that based on... well, your assumptions about the market as a whole. a few dots ahead, you point that the enginering improvement is simple, easy, well known and - yes - expensive. Comparatively, it is as expensive as purchasing, from the other side of the globe, a single unit of a very niche non consumer grade motherboard... and we know is not absurd, since we do this all the time, here in this same forum.

As a matter of fact, the very existence of HHHL GPUs is a proof that the market dont care so much about price, since these GPUs sit a bit behind its non-HHHL counterparts, considering raw performance and thermals.
Marketing statistics? Where? My opinion is based on an assessment of production cost, sale price, value and performance.

You're also dismissing some very serious concerns as "simple, easy". Sure, making a PCB with more layers is by itself a simple process. Designing such a PCB with overlapping VRAM traces without causing interference is in no way trivial. Nor do you seem to understand the effect this has on cost - it is significant. An example: look at AM4 ITX boards vs. Intel-platform ITX boards. The cheapest AM4 ITX boards are about 50% more expensive than the cheapest Intel-compatible ones, and this is largely due to AM4's requirement for a thicker PCB. This was also cited by OEMs as the reason for AM4 ITX boards being slow to market. The added cost and complexity is very significant, and will still be so even for a small board like a HHHL GPU. You dismiss this as "just buy a thicker board", which makes it seem like you don't understand the design requirements for a component like this whatsoever. Designing them to comply with signaling specs for sensitive components like RAM is no trivial matter. This exact issue contributes a very large portion of the added cost for HEDT motherboards too, which have the same issue with having to squeeze more RAM traces into a largely similar area as cheaper boards (though they have more space and don't have to resort to exotic solutions like rear-mounting significant components).

Also you IMO seriously underestimate the effect of price and value on sales. Sure, SFF enthusiasts are willing to pay more for more in a small space. But where does this willingness end? A 1050 LP costs and performs roughly the same as a normal 1050. A 1060 squeezed into the same space, or the 1660 I mocked up above will have dramatically less performance than a stock card (no more than 4 memory channels means 33% less memory bandwidth, which means at the very least 20-30% less performance), while costing the same or more (any savings from less VRAM will be eaten by the added board complexity and making a sufficient cooler in the small space). The lack of room for a proper VRM setup would likely impose an artificial power limit/lower boost clocks, or at the very least make overclocking impossible. Would you be willing to pay $230 for a significantly crippled 1660? I'm sure some would, but the vast majority would go for a slightly larger case (still plenty of SFF cases fitting non-LP GPUs) and get the extra 20-30% of performance for the same money.

Tl;dr: IMO you're underestimating the cost, complexity and performance impact of this, while overestimating the willingness for people to pay significant premiums for SFF performance.
 

cleveland

Master of Cramming
Sep 8, 2016
455
240
Tl;dr: IMO you're underestimating the cost, complexity and performance impact of this, while overestimating the willingness for people to pay significant premiums for SFF performance.

Thanks for the TL;DR. I'm now inclined to believe that you've changed your mind about that part where you stated it was not possible to create such a GPU with just the tech we have today.
 

Valantar

Shrink Ray Wielder
Jan 20, 2018
2,201
2,225
Thanks for the TL;DR. I'm now inclined to believe that you've changed your mind about that part where you stated it was not possible to create such a GPU with just the tech we have today.
Well, I didn't say that, so I guess a clarification is in order. I said a full-featured TU116 (or was it a 1060 die? Doesn't matter much either way.) can't be fit onto a HHHL card. I still stand by that, as there are two options as I see it: ditch two memory channels outright (no longer full featured), or attempt to mount two memory chips on the back of the card (might be possible, but is highly unlikely due to the sensitivity of memory signaling, no matter how thick you make the PCB, let alone direct physical limitations such as where on the package the traces originate, which determines RAM chip placement to a large degree). That still leaves out the issue of fitting a 120W VRM and all ancillary components and then cooling all of this, neither of which are trivial matters in that form factor. The VRM might need more space than the board has available due to the large package and PCIe power connector.

I think fitting a mid-sized chip with HBM in a HHHL form factor would be feasible (not quite R9 Nano sized, but 2-300mm2?), but that wouldn't be a TU116. So no, I haven't changed my mind.
 

Thehack

Spatial Philosopher
Creator
Mar 6, 2016
2,812
3,670
J-hackcompany.com
I don't think anyone here is saying it's completely impossible.

IMO, saying use "better engineering skills" is quite dismissive and shows a certain misunderstanding of the purpose of engineering. Good engineering purpose is to create feasible, marketable, and usable product that meets the demands of the customer. There is no point to use "better engineering" when the end product is a unicorn GPU that no one is willing pay for.

Calling it a "thicker plate" and requiring "better engineering" is just being ignorant and impractical. Engineering traces, verification, certification, marketing, packaging, production, it's not just engineering in the equation.

At this point, and costs, it would be far simpler to just build an microSTX system and buy an MXM GPU. HHHL only exists due to business desktop that need light graphics work. That's what they're designed to sell to.
 
Last edited:

Valantar

Shrink Ray Wielder
Jan 20, 2018
2,201
2,225
Reviving this thread, anyone know if we can expect something new from AMD?
Sadly they've been silent on the low end front since the 5500 XT launched, and I don't think there have even been laptops with the mobile 5300 they announced simultaneously yet. Given that the 5500 XT is a ~130W part I am really hoping for a lower end SKU to show up. Though if rumors of RDNA2 having significant efficiency improvements turn out to be true I certainly wouldn't mind waiting for something based on that. It might also be that AMD isn't prioritizing this due to low 7nm wafer availability, holding off on making even more GPU silicon until they have more flexibility, which should be clearing up in the coming months as Apple moves on to 5nm. They have a lot of spaces in the GPU lineup to fill still from <100W low-end/LP to high end and flagship level, so I think it'll be some time until we see a fully fleshed out RDNA lineup. And typically the lowest end cards are the last to launch, sadly.
 
  • Like
Reactions: Jello

Analogue Blacksheep

King of Cable Management
Original poster
Dec 2, 2018
849
705
Bringing this thread back because we have a new GPU generations from AMD and Nvidia that might bring up some interesting dicussion about Low Profile Cards.

Think we will see triple slot Low Profile cards this generation? Hell, will we even see Low Profile cards appear in the future. If we do I doubt it will be until late 2021/early 2022.
 

robbee

King of Cable Management
n3rdware
Bronze Supporter
Sep 24, 2016
881
1,387
Think we will see triple slot Low Profile cards this generation? Hell, will we even see Low Profile cards appear in the future. If we do I doubt it will be until late 2021/early 2022.

With power usage being increased so much in the current generation, I don't see a low profile GPU any time soon. The TDP's have just increased too much:

1080: 180w
2080: 245w
3080: 320w

1070: 150w
2070: 175w
3070: 220w

1060ti: 120w
1660ti: 120w
3060ti: 200w

1050: 75w
1650: 75w
3050 (?): if anything is clear from the above, it will probably not be 75w

We might see something like a RTX 3030 or RX 6300 but I doubt it's even near to being the focus of Nvidia or AMD at the moment...
 
  • Like
Reactions: Bubba

GuilleAcoustic

Chief Procrastination Officer
SFFn Staff
LOSIAS
Jun 29, 2015
2,984
4,421
guilleacoustic.wordpress.com
Asus announced RTX 3060Ti Dual Mini cards in their 3060Ti lineup today. Excited for that.

Not really low-profile ....

 

Valantar

Shrink Ray Wielder
Jan 20, 2018
2,201
2,225
Asus announced RTX 3060Ti Dual Mini cards in their 3060Ti lineup today. Excited for that.
Not really low-profile ....

There's also a really tiny PCB one from PNY.

Bringing this thread back because we have a new GPU generations from AMD and Nvidia that might bring up some interesting dicussion about Low Profile Cards.

Think we will see triple slot Low Profile cards this generation? Hell, will we even see Low Profile cards appear in the future. If we do I doubt it will be until late 2021/early 2022.
With power usage being increased so much in the current generation, I don't see a low profile GPU any time soon. The TDP's have just increased too much:

1080: 180w
2080: 245w
3080: 320w

1070: 150w
2070: 175w
3070: 220w

1060ti: 120w
1660ti: 120w
3060ti: 200w

1050: 75w
1650: 75w
3050 (?): if anything is clear from the above, it will probably not be 75w

We might see something like a RTX 3030 or RX 6300 but I doubt it's even near to being the focus of Nvidia or AMD at the moment...
We might, but with the continually expanding span of performance in a generation it'll likely be RTX(?) 3040 or similar, as the 3050 is likely to step up a bit in power draw. Given just how much faster the current 80 and 90 tier cards are, there's a need for more product tiers to fill out the power and price stack. I certainly wouldn't mind a 75W RTX 3040. AMD is likely to do better in that segment given their current efficiency lead - a 75W RX 6500 or 6400 XT would be a very attractive proposition, particularly if it kept some Infinity Cache - that would make up for the narrow memory bandwidth forced by the LP PCB.
 

Analogue Blacksheep

King of Cable Management
Original poster
Dec 2, 2018
849
705
@SFFManiac07 While it is not an LP card, I do like the look of that RTX 3060Ti Dual Mini. Means we are getting new ITX options which is a good sign.

@robbee @Valantar - Current rumours/speculation are pointing to the 3050 to being 90W. and Tech Power Up is guestimating 150W for the 3050 ti. That said, a 3040 would be cool.

I would be curious to see if someone experimented with copper similar to the amount ASUS are using for the 3090 blower card they annouced today.

Also, who's to say it might come from the usual suspects. ASL came out with this single slot 1650, so maybe they could try something (and a dealer in the West would be great).
 

Valantar

Shrink Ray Wielder
Jan 20, 2018
2,201
2,225
@SFFManiac07 While it is not an LP card, I do like the look of that RTX 3060Ti Dual Mini. Means we are getting new ITX options which is a good sign.

@robbee @Valantar - Current rumours/speculation are pointing to the 3050 to being 90W. and Tech Power Up is guestimating 150W for the 3050 ti. That said, a 3040 would be cool.

I would be curious to see if someone experimented with copper similar to the amount ASUS are using for the 3090 blower card they annouced today.

Also, who's to say it might come from the usual suspects. ASL came out with this single slot 1650, so maybe they could try something (and a dealer in the West would be great).
I certainly wouldn't be surprised by a 150W 3050Ti - we need to remember that the 16xx GPUs were a tier below their 20xx cousins when comparing the 3rd digit of the name - i.e. the 1660/- Ti were a tier below the 2060. If there's no GTX 26xx GPUs this time around, that necessitates the addition of at least one more tier, as a 3050/- Ti would then be a successor to the 1660/- Ti, not the 1650. So with the 3070 at 220W and the 3060 Ti at 200W (rated), I would guess 175W for a 3060, 150 for a 3050 Ti and 125-130 for a 3050 would make sense, though I guess they could go lower. Below that is pretty much an open question, as that's the point where you need to start making hard feature cuts to cut power noticeably - less memory channels and so on. 3040 Ti at 100W and 3040 at 75W could happen, though I wouldn't be surprised if they skipped the 100W level - there's such a thing as too many Ti models, after all. But 150W for a 3050 Ti and 90W for a 3050 doesn't make sense to me - that's a 66% power increase in a single product tier. That's too high.

That 3090 blower heatsink is certainly impressive, though given how densely packed HHHL GPU PCBs are I doubt we'll see a large vapor chamber - it would be near impossible for it to clear the VRM components and the like. But even just a CPU+VRAM vapor chamber with a dense chunk of copper on it would make for a very interesting GPU - maybe a 125W model with an XT90 power connector and a bundled adapter? I'd be into that, for sure.

I doubt we'll see any >2 slot LP GPUs simply because the number of cases able to fit them would be pretty close to zero.
 

Snerual

Cable-Tie Ninja
Jul 3, 2020
196
224
I doubt we'll see any >2 slot LP GPUs simply because the number of cases able to fit them would be pretty close to zero.
My SKTC A02 could easily fit 3 slot assuming there's only 2 physical brackets on the card.

Silverstone ML09 and ML06E could do 2,5 slot, though I'd prefer to stick to my "deshroud + 2 80x25mm fans squeezed between heatsink and side panel" set up that makes the card super quiet.
 
  • Like
Reactions: SFFManiac07