Custom Graphics Card-theoretical question.

[John_Rich]

So, 

 

I was hoping that someone with more insight could answer this.  Why are Graphics cards only sold as full units?  

 

I have 2 reasons for this question: 

1. If the end user is going to remove the stock cooler and add a watercooler on, then you could grab cards out of your normal production line before they get a cooler added and sell them for basically the same price.  You would be decreasing your costs (don't have to buy a cooler for it) and therefore increase your profit, and you'd end up making the water cooling easier on the enthusiast, by supplying a naked pcb.

2. Because the Graphics card is basically a mini-computer in its own right.  Why can't we get a board that can receive a GPU and video RAM like a MB receives RAM and a CPU?  There seems to be an under served market for Graphics card customizations.  For example, maybe your GPU has plenty of CUs but you need more VRAM.  Right now you have to do a whole new card, which is more expensive than a couple sticks of VRAM.

 

It just seems like this is something that should exist, so why doesn't it?

[TudorMN]

Maybe is exclusive for high demands only not for small order?

 

[mariushm]

The memory has to be soldered onto the circuit board of the video card because any kind of connector would add extra trace length and resistance and can cause issues due to poor contacts between memory sticks and connector pins.

There's hundreds of traces (wires) between the ram chips and the gpu chip and all those traces have to be length matched to sub-millimeter lengths

 

Also, a memory stick in the computer has a 64 bit bus while video cards are often 192-256 bit so you'd basically have to install 4, 6 or 8 sticks of video card memory onto a video card ... where would you fit so many connectors and think of how all those traces between the gpu and the memory slots would have to work around the other slots.

 

As for cooling, there are water blocks for video cards, and they try to keep the footprint (the screw holes around the gpu chip) at the same distances to make blocks compatible. between models of video cards with same gpu chip

 

Coolers on video cards have more purpose than just cooling the gpu chip, they also have to cool the vrm (the dc-dc converter which converts 12v to the voltages the gpu chip needs, and the vrm which produces voltage for memory chips) and the cooler often cools the memory chips as well.

 

[Himommies]

12 minutes ago, John_Rich said:

Why are Graphics cards only sold as full units?  

I imagine it has to do with QC and economies of scale. It's probably still cheaper to make 1000 GPU's with coolers than having 2 assembly lines and producing 990 gpu's with coolers and 10 without. Also QC would be really terrible since you basically have to waste employee time mounting a GPU cooler to test it out.

[GoldenLag]

23 minutes ago, John_Rich said:

You would be decreasing your costs (don't have to buy a cooler for it) and therefore increase your profit, and you'd end up making the water cooling easier on the enthusiast, by supplying a naked pcb

its a very niche market of a allready niche market. 

 

they are better off making a PCB on their reference card that isnt trash and just have people who want to watercool, buy the reference card. 

 

edit: and for the Vram thing, its physically improbable to do on higher end cards. its allready borderline doing 384-bit busses on a PCIe card around a GPU while its all on one PCB

[DildorTheDecent]

27 minutes ago, John_Rich said:

1. If the end user is going to remove the stock cooler and add a watercooler on, then you could grab cards out of your normal production line before they get a cooler added and sell them for basically the same price.  You would be decreasing your costs (don't have to buy a cooler for it) and therefore increase your profit, and you'd end up making the water cooling easier on the enthusiast, by supplying a naked pcb.

 

If you're buying a Nvidia GPU this goes against the greenlight programme.

 

They dictate that all GPUs must ship with a cooler. They (the board partners) cannot supply naked PCBs for the mass market/purchase. There was a debate about this during the Kepler era. After the GTX 680 had came to market, newer 2nd gen (GTX 700) cards faced some limitations. Factory voltage control was also limited quite a bit too.

 

I'm not sure how AMD regulates there market, but I'd imagine that it's pretty similar.

[John_Rich]

14 minutes ago, mariushm said:

Coolers on video cards have more purpose than just cooling the gpu chip, they also have to cool the vrm (the dc-dc converter which converts 12v to the voltages the gpu chip needs, and the vrm which produces voltage for memory chips) and the cooler often cools the memory chips as well.

 

Of course they do, but that's the same issue as when a cooler is removed and a water cooler is added.  

 

As far as the soldering of memory being required for a good connection, desktop CPUs aren't soldered in, and that seems like a much more serious juncture.  I'm sure that some kind of "proper pressure" mounting system, like the AMD TR CPUs could be made to ensure that the VRAM was properly seated.

 

19 minutes ago, Himommies said:

I imagine it has to do with QC and economies of scale. It's probably still cheaper to make 1000 GPU's with coolers than having 2 assembly lines and producing 990 gpu's with coolers and 10 without. Also QC would be really terrible since you basically have to waste employee time mounting a GPU cooler to test it out.

I can't imagine that the cards are not tested before you add a cooler, it would just be far too wasteful to either trash a perfectly good cooler, or recover it.  Therefore, you could just grab a stack of 100 cards without coolers between these steps.

 

9 minutes ago, GoldenLag said:

its a very niche market of a allready niche market. 

 

they are better off making a PCB on their reference card that isnt trash and just have people who want to watercool, buy the reference card. 

 

I think that you may have the real answer here, perhaps it is just too niche.

[John_Rich]

1 minute ago, DildorTheDecent said:

If you're buying a Nvidia GPU this goes against the greenlight programme.

 

They dictate that all GPUs must ship with a cooler. They (the board partners) cannot supply naked PCBs for the mass market/purchase. There was a debate about this during the Kepler era. After the GTX 680 had came to market, newer 2nd gen (GTX 700) cards faced some limitations. Factory voltage control was also limited quite a bit too.

 

I'm not sure how AMD regulates there market, but I'd imagine that it's pretty similar.

Ahh, that sounds more like something that would cause this.  I wonder if an indie company could do this?

[Jurrunio]

1. It's Nvidia and AMD's policies that requires every graphics card sold to be a full unit. Back in the past (some time in the mid 2000s) there's basically no regulation, as a result all sorts of cut down cards in power delivery, even memory bus widths are available. These policies came out to clean up the big mess in the past. There are still occassionally cards that are clearly meant to do what you suggest, i.e. the MSI 1080ti Armor with the same cooler you'd find on the GTX 1060, but PCB of the 1080Ti Gaming X.

 

2. 

50 minutes ago, John_Rich said:

Because the Graphics card is basically a mini-computer in its own right.

No it's not, a graphics card is too dedicated for a couple of tasks to be considered a computer. It's only a component.

 

50 minutes ago, John_Rich said:

Why can't we get a board that can receive a GPU and video RAM like a MB receives RAM and a CPU? 

because it's cheaper to manufacture more of the same stuff, less debug time and trouble of updates for example. Look at how CPU, motherboard and memory struggle in compatibility, especially on Ryzen. If all cards carry the same specs, then all that matters (in terms of compatiblity) is the connection to the system. 

 

50 minutes ago, John_Rich said:

For example, maybe your GPU has plenty of CUs but you need more VRAM.  Right now you have to do a whole new card, which is more expensive than a couple sticks of VRAM.

Graphics cards don't face this problem anywhere near as much as CPU and system memory, mostly because GPU is a type of coprocessor with a very limited range of tasks it's expected to perform. CPUs on the other hand have to cope with everything there's no dedicated hardware for, so the focus on being good at everything increases the need to be more customizable.

[GoldenLag]

15 minutes ago, John_Rich said:

As far as the soldering of memory being required for a good connection, desktop CPUs aren't soldered in, and that seems like a much more serious juncture.  I'm sure that some kind of "proper pressure" mounting system, like the AMD TR CPUs could be made to ensure that the VRAM was properly seated.

now imagining making 12 of those sockets around a GPU die that will be the size of about a 7700k die. all of the sockets will have to be BGA or LGA, i believe it uses BGA. 

 

then you need to add mounting system for all of those. you have allready massivly increased your cost making that PCB. and now you need to sell memmory chips that people can buy. 

 

and the issue isnt as much to do with the only the contact, but also the tracing. also each socket will need to be added after the PCB is done, which makes it really expencive. 

 

also you can nolonger make a 384-bit buss on GDDR6, which rules out the 2080ti, 1080ti, Titan p, r9 290, r9 290x, r9 390, r9 390x, gtx 980ti, gtx 780ti. just because its allready hard enough fitting the dies without making sockets for them. 

 

i can imagine gddr5 socketed system for 384-bit maybe working. but it steals so much PCB space. 

 

 

edit: 

worst case scenario for reference (512-bit bus layout). 

 

 

not worst case scenario, but stil bad (384 bit layout):

[DildorTheDecent]

13 minutes ago, John_Rich said:

Ahh, that sounds more like something that would cause this.  I wonder if an indie company could do this?

If that indie company makes there own GPUs then they can do as they please. The only rules they would need to abide by is their own and the country they're selling in.

 

However, if they are using (in my example) nvidia GPUs, then they must abide by nvidia's rules if they wish to sell to the mass market. Nvidia (to my knowledge) approves every "variation" of there GPUs. So a HOF, K|NGP|N, Matrix must be approved by them before it is sold to market.

 

Those requirements (again, to my knowledge) are:

1. The GPU must have a cooling solution
2. The GPU must have limitations voltage control
   1. There is nothing to stop you using a tool that can talk to the voltage controller and request everting however. It just has to not "come with it" if you catch my drift.

 

They (the board partners) could hand out/sell cards to extreme overclockers since they're just going to use a LN2 pot anyway but that market is very, very small.

[Mihle]

GPUs are more memory sensitive than what CPUs is. There are much more traces between a GPU and it's memory than a CPU and it's memory.

[John_Rich]

19 hours ago, Jurrunio said:

No it's not, a graphics card is too dedicated for a couple of tasks to be considered a computer. It's only a component.

So, it's like a mini computer that is specialized to a single task and not a general all purpose thing, I would probably just say "mini computer" for short and expect people to understand that it's an approximation, not an exact statement.

 

@GoldenLag

You have a good point there, but honestly looking at those cards I was thinking more of a "how could it be done" kind of thing.  I still think that it's possible, and could even give a gain to enthusiasts, however I do see why it doesn't happen.

 

@DildorTheDecent

Thanks!