Why dont video cards have upgradable memory?

|oR|Spetsnaz

2[H]4U
Joined
Jul 20, 2004
Messages
2,797
With cards getting into the 256-512MB range, it seems that it would make sense to start equpping them with ram slots for upgrading. It would be nice if you could get a 6800 128mb like I have and stick in another chip with 128-256mb of GDDR3 on it if you feel the need to. Are video card companies not letting us do this because it would cut back on their profits? I think when you plunk down $500-700 on a top of the line card, it should have some upgradability.

What are your thoughts?
 
Not really needed nowadays.

Look at the x800xl 256mb vs 512mb card. They perform almost exactly the same. It's the core that makes the video card.
 
IceDigger said:
Not really needed nowadays.

Look at the x800xl 256mb vs 512mb card. They perform almost exactly the same. It's the core that makes the video card.
QFT

On the other side of things NV is going to be using TurboCache on their IGP NF4 boards so in a way you're going to be getting the "upgradeable video card" that your asking for. ;)
 
I think it would be cool if they had PCI-E slots for upgrading. Then you just put another video card on that PCI-E slot. Then that next one can have two PCI-E slots, one on each "face" of the card. Then you can attatch two more to that one. That way you'll have a cluster of 4 video cards. Then SLI a matched pair of that and you'll break 70,000 3dmarks :cool:
 
Because they can make you buy an entire new card if they aren't upgradable.
 
Old ISA and PCI cards and even some first generation AGP cards (e.g. ATI Rage) had slots to upgrade memory. I think at today's clock rates, signal quality is so sensitive that an extra connector and even slightly longer traces would not work at all.

Besides that, by the time you actually need that extra bit of video memory it's likely that your card will be too slow to run anything that makes full use of it.
 
Flo said:
Old ISA and PCI cards and even some first generation AGP cards (e.g. ATI Rage) had slots to upgrade memory. I think at today's clock rates, signal quality is so sensitive that an extra connector and even slightly longer traces would not work at all.

Besides that, by the time you actually need that extra bit of video memory it's likely that your card will be too slow to run anything that makes full use of it.

Signaling might be part of the reason. Personally I think its cost. When we did have upgradable video memory, rarely did anyone take advantage of that feature. I did, but I don't know if I know anyone else who ever did.
 
Flo said:
Besides that, by the time you actually need that extra bit of video memory it's likely that your card will be too slow to run anything that makes full use of it.

Not entirely true. All the new cards came out last year and Doom3 was unplayable at ultra settings unless you had 512mb of GDDR3. Was my card already too old ?

It's probably a lot better using ram that's pre-installed on the card rather than upgrading. How much are you gonna upgrade ? Faster ram requires a whole new card since we're not talkin about normal system ram. Larger amounts is kinda pointless because you don't really need to upgrade the ram by another 256MB and this will cut sales of newer video cards.

They simply want you to buy a new card instead of buying ram sticks that might not be compatible. Plus it would make your card look ugly :p
 
All the new cards came out last year and Doom3 was unplayable at ultra settings unless you had 512mb of GDDR3. Was my card already too old ?
Doom 3 is a bad example as ultra quality looks barely better than high. I don't consider Doom 3 a game that makes good use of the extra memory, but ymmv.
 
Flo is right on this one. From an engineering standpoint it isn't feasible. Signal quality, and ground bouncing would be too hard to control. You would be increasing your trace lengths, and adding an external socket through which all of the signals would need to go. The memory speeds we are seeing now just wouldn't be possible if the memory were socketed like that.

I also wouldn't forget that you are looking at space problems. Heatsinks on current products already cover almost the entire GPU/VPU side of the card. Even if you used a low profile socket of some kind and install the RAM on the backside of the card, you would still risk having space issues with the CPU, or more likely, the motherboard's north bridge chipset on some motherboards.
 
Cooling the ram is not an issue .. you can use bga ram sinks. The upgradibility is not really worth the trouble, he is right. By the time you need a ram upgrade, your card is most probably obsolete.
 
It would be a major headache for manufacturers to have an extra variable they can't control in the manufacturing testing phase and hence uneconomical.

If you "openware" the ram slots, then there will be possible compatibility issues, etc, whereas in the case where only the manuf can make the chips... just seems like an extra headache :)
 
f U z ! o N said:
heck lets go all out and make a vid card with upgradeable core and mem. that would be cool.

While that will never happen, a base PCB that you could change GPU's and memory on would be sweet. I'd like to buy a plain Nvidia or ATi card and the GPU of my choosing later. Add the memory. It would be like a mini-motherboard.

It would likely be larger, but cool none the less.
 
aranaxon said:
It would be a major headache for manufacturers to have an extra variable they can't control in the manufacturing testing phase and hence uneconomical.

If you "openware" the ram slots, then there will be possible compatibility issues, etc, whereas in the case where only the manuf can make the chips... just seems like an extra headache :)

Not only that, but could you imagine the support calls related to end-users accidentally putting in DDR memory in place of GDDR3? You'd have to ensure a variety of different packaging types to prevent this instead of just little surface mount chips.

The design costs would soar and the companies would need to ramp up their customer support desks for the significant number of calls for this.

My cards would increase in price a lot for a feature I may not use. I'm not excited by this prospect and see the reason this hasn't been performed. Heck, any single OEM maker could have done this, but saw that the benefits were not better than the extra headaches involved.
 
Sir-Fragalot said:
While that will never happen, a base PCB that you could change GPU's and memory on would be sweet. I'd like to buy a plain Nvidia or ATi card and the GPU of my choosing later. Add the memory. It would be like a mini-motherboard.

It would likely be larger, but cool none the less.
It might look cool, yes, but I doubt that it'll be more efficient than having a single, optimized set of components and layout.

I very much doubt that it'd be any cheaper than just buying a new card every couple of years as well, as production runs of GPUs are often very short, not to mention the added costs of the special types of RAM these GPUs use.
 
Flo said:
Old ISA and PCI cards and even some first generation AGP cards (e.g. ATI Rage) had slots to upgrade memory. I think at today's clock rates, signal quality is so sensitive that an extra connector and even slightly longer traces would not work at all.

Besides that, by the time you actually need that extra bit of video memory it's likely that your card will be too slow to run anything that makes full use of it.

Un4given said:
Flo is right on this one. From an engineering standpoint it isn't feasible. Signal quality, and ground bouncing would be too hard to control. You would be increasing your trace lengths, and adding an external socket through which all of the signals would need to go. The memory speeds we are seeing now just wouldn't be possible if the memory were socketed like that.

I also wouldn't forget that you are looking at space problems. Heatsinks on current products already cover almost the entire GPU/VPU side of the card. Even if you used a low profile socket of some kind and install the RAM on the backside of the card, you would still risk having space issues with the CPU, or more likely, the motherboard's north bridge chipset on some motherboards.

QFT. It would be a step backwards (remember BGA? it was pretty important)
 
Elledan said:
It might look cool, yes, but I doubt that it'll be more efficient than having a single, optimized set of components and layout.

I very much doubt that it'd be any cheaper than just buying a new card every couple of years as well, as production runs of GPUs are often very short, not to mention the added costs of the special types of RAM these GPUs use.

Well I didn't get into the practicality of it. It would be cool though IMO.
 
Actually, I had a videocard once that had upgrade-able memory. It was sweet. I dont even remember the brand of it though. It was a very large card too, we are talking old. I'll see if I can find it to take a pic.
 
omg w00t! I found it! couldnt find any of the little ram chips though. The thing has a whopping 2mb of onboard ram and you can add up to 2 more sticks!!!


my.php
 
Sir-Fragalot said:
While that will never happen, a base PCB that you could change GPU's and memory on would be sweet. I'd like to buy a plain Nvidia or ATi card and the GPU of my choosing later. Add the memory. It would be like a mini-motherboard.

It would likely be larger, but cool none the less.

I seem to remember there WAS a company talking about doing this, though I think it refered more to integrated video than add-in cards. This was back in ~1997, I think. Can't remember who it was though. There were a few more companies around back then...


edit: lol, was searching for something about it and found this...
About older cards...

SOME OLDER CARDS MAY NOT BE ABLE TO HANDLE some of the newer "game" style programs in our catalog. You cannot tell if that card can run EVERY program until you try it with every program.

Litmus Test:
Among our software programs, I often suggest trying to run Joseph's Story CD or Bongo Loves the Bible CD to test your computer's capabilities. If they run smoothly on your computer, it probably means that you have a good minimum about of processing MHZ, ram and a good videocard that can run just about everything else in our catalog. (I'm writing this comment in 2005.)
Can we get these benchmarks added to [H] reviews please ;)



u_95244
 
Yeah, with memory running at 1+ Ghz, I can understand why things like the length of the traces and signal quality would come into play. None the less, its a shame that what can be the most expensive part of your system is not upgradable. One of the reasons I build my own computers and stick to PC's over consoles is because I can upgrade them over time. It sucks when you plop down $500-700 and know that your card is going to lose a sizable portion of its value down the road and there's nothing you can do but plop down another $500-700. I guess thats PC's for ya...
 
low latency ram, high speeds found on todays cards require precise trace lengths, hence the odd memory layout on some of the new cards..... any kind of upgrade socket would lie way too far away from the GPU.....
 
I'm not sure if signal quality is the reason. After all, main memory is also very fast and that still works through DIMM slots. It may be more expensive to design, but I don't think it's impossible.
Judging from the prices of videocards, it seems that the price is almost solely determined by the amount and speed of memory.
That may have something to do with it... There are too many types of memory used on the videocards, so it would be expensive to sell them separately (demand for each particular type would be relatively low). Also, by the time you upgrade, there is probably a new generation out already (they tend to do that every 6 months)... Which makes it all even harder and more expensive.
I guess it's just cheaper to not bother with upgrading at all and just buy a new card.
 
Scali said:
I'm not sure if signal quality is the reason. After all, main memory is also very fast and that still works through DIMM slots. It may be more expensive to design, but I don't think it's impossible.
Judging from the prices of videocards, it seems that the price is almost solely determined by the amount and speed of memory.
That may have something to do with it... There are too many types of memory used on the videocards, so it would be expensive to sell them separately (demand for each particular type would be relatively low). Also, by the time you upgrade, there is probably a new generation out already (they tend to do that every 6 months)... Which makes it all even harder and more expensive.
I guess it's just cheaper to not bother with upgrading at all and just buy a new card.

Try and buy some 1.6 or even 2.0ns Samsung GDDR3 Ram on a dimm....... Plus what alot of people are missing is the fact that the Companys actaully can make money on the ram. Better ram does not instantly double the value of the card but look at the price difference between a DDR256 meg card and a GDDR 256 meg card.
 
Scali said:
I'm not sure if signal quality is the reason. After all, main memory is also very fast and that still works through DIMM slots. It may be more expensive to design, but I don't think it's impossible.
I'm willing to bet that speeds and latency of main memory would improve considerably if it were soldered directly to the mainboard's PCB.

Just look at the average connector: its conducting surfaces are never as clean as they look. On a microscopic level you'll see that they're neither smooth (re lapping heatsinks) or clean (corrosion, oxidation, other impurities, etc.). Adding a layer of gold or similar helps to some extent, but nothing beats a single, uninterrupted wire or trace.

And main memory is not fast. It's horribly slow. A reasonably fast CPU has to wait hundreds of cycles before data from main memory arrives (an on-die memory controller nearly halves this latency). Even the on-die L2-cache is relatively slow as it is.
 
Elledan said:
I'm willing to bet that speeds and latency of main memory would improve considerably if it were soldered directly to the mainboard's PCB.

Latency is no issue for a deeply pipelined architecture such as a 3d accelerator. GDDR actually has much larger latencies than regular DDR.

And main memory is not fast. It's horribly slow. A reasonably fast CPU has to wait hundreds of cycles before data from main memory arrives (an on-die memory controller nearly halves this latency). Even the on-die L2-cache is relatively slow as it is.

Again, that's latency, which is not relevant. You should look solely at the bandwidth.
 
Scali said:
Latency is no issue for a deeply pipelined architecture such as a 3d accelerator. GDDR actually has much larger latencies than regular DDR.



Again, that's latency, which is not relevant. You should look solely at the bandwidth.
Methinks you are talking about videocards, whereas I was talking about CPUs and mainboards :)

Remember RDRAM? Crazy speeds and bandwidth, but also crazy latency (and heat). Bandwidth isn't everything.
 
Elledan said:
Methinks you are talking about videocards, whereas I was talking about CPUs and mainboards :)

Silly me, I thought this thread was about video memory.
 
His point about mainboard memory was relevant to the thread.

I'd really be interested to know if anybody like ATI or nVidia has been toying around with the idea of upgradable memory. I think a lot of people would be more apt to purchase a $700 video card if they knew it was upgradable. Like if you could buy a single GPU 6800U w/ 256MB and have an extra socket (like a CPU socket) where you could install another GPU and a low profile, angled memory expansion slot down the road if you felt so inclined.
 
|oR|Spetsnaz said:
I'd really be interested to know if anybody like ATI or nVidia has been toying around with the idea of upgradable memory.

Judging from this interview, I would say ATI isn't, anyway..

Q: Why doesn’t ATI create a graphics card with an upgradeable VPU? I know the Flip-Chip gives me such ideas like upgrading CPU’s.

A: We don’t create ‘socketed’ boards, because the use of a socket would significantly reduce the quality of the signals used to communicate with the memory chips. We would need to run slower to do that. That’s not acceptable.

from http://www.pcgameworld.com/article.php/id/151/

u_95244
 
While that will never happen, a base PCB that you could change GPU's and memory on would be sweet. I'd like to buy a plain Nvidia or ATi card and the GPU of my choosing later. Add the memory. It would be like a mini-motherboard
Heh...I thought of this before, and agree it would be nice. :D
I'd pay to pop a new GPU on mine.
 
Monkey34 said:
Heh...I thought of this before, and agree it would be nice. :D
I'd pay to pop a new GPU on mine.

Sorry, my thought kindof progressed from yours and somebody else's comment about how specialized graphics card's PCB layout can be. I was thinking of something like the cards with dual 6800GT chips on them, except with one of them being hard wired in and the other one would be left out and have a CPU socket type assembly in its place for future upgrading. But as evidenced by that post with the ATI interview, I guess its just not possible with the speeds they are wishing to achieve. Quite a shame if you ask me.
 
Problem with GPUs is that next-gen GPUs also tend to come with next-gen memory. So if you don't replace the memory, it will become horribly bandwidth-limited.
And then ofcourse videocards change memory systems more often than some people change underwear, so it'd be hell to make GPUs that are compatible with old and new boards.
I suppose videocards are a typical device where it's really all or nothing.
 
|oR|Spetsnaz said:
With cards getting into the 256-512MB range, it seems that it would make sense to start equpping them with ram slots for upgrading. It would be nice if you could get a 6800 128mb like I have and stick in another chip with 128-256mb of GDDR3 on it if you feel the need to. Are video card companies not letting us do this because it would cut back on their profits? I think when you plunk down $500-700 on a top of the line card, it should have some upgradability.

What are your thoughts?

Its all about the money.

The ONLY reason it was ever an option back in the 90's was when memory cost was $30-$40 for one megabyte. A card could be $60-$90 cheaper with only two MB ram while an upgrade option to 4 MB ram by inserting memory chips was still an option.

I think I still have some old Matrox and S3 cards with open memory sockets in them.
 
Crapgame said:
Because they can make you buy an entire new card if they aren't upgradable.

You mean the video card companies want to keep making money ?

EXACTLY...

From above post...modified ;)

"A: We don’t create ‘socketed’ boards, because the use of a socket would significantly reduce profits used to communicate with the CFO's in the company. We would need to run about the same to do that. That’s not acceptable."
 
Back
Top