VRAM Upgrade?

[jayhughes79]

Can anyone tell me if this is a silly idea?

Why isn't there a graphics card that has VRAM slots like a motherboard has RAM slots so we can by extra/upgrade if we want?

 

[tommybhoy]

If gpu manufacturers simply drip feed vram amounts, they make way more profit persuading you to buy the next big thing.

 

[jayhughes79]

Damn them! *shakes fist*

 

[RanxZy]

Yup it's the limit being raised just when new games are starting to push the barrier that keeps us coming back for more.

 

[RatmanGB]

There used to be graphics cards that allowed this.

I had a PCI ATI Mach-64 that I upgraded from 1Mb to 2Mb and a Matrox G200 that I upgraded from 4Mb to 8Mb back in prehistoric times!

 

[ubersonic]

It's been done, first they offered extra sockets for the chips:

Then they put the chips on expansion cards:

---------

Personally I always thought it would be cool if Nvidia brought out a standalone PhysX card that could also share it's VRAM with the main GPU.

 

[Meaker]

The reason is that those cards are not running an 8Ghz memory bus with all the signal integrity issues that that brings.

 

[ubersonic]

The reason is that those cards are not running an 8Ghz memory bus with all the signal integrity issues that that brings.

It's actually more to do with cost. Mantle, Direct3D 12 and OpenGL Next will all allow VRAM stacking in multi card setups (I.E two 4GB cards in SLi/CF can work as 8GB not mirrored 4GB). If you can do it via two PCI-E slots and a connecting bus then you can definitely do it via a DIMM slot.

And when I say cost I mean partly the cost to implement, and also the cost of users upgrading their GPUs instead of upgrading to your newer models.

 

[tres_kun]

Imagine Vram being faster than ram by magnitudes x faster
X unknown
Imagine the pin connection being larger than a DDR3
Stick that on a GPU
Question will be how large will a GPU have to be to accommodate a socket
And the size of the socket to accommodate the higher bandwidth requirement

 

[ubersonic]

And the size of the socket to accommodate the higher bandwidth requirement

At a guess, about the same size as a x4 PCI-E slot (minus the power section).

 

[Kaapstad]

It's actually more to do with cost. Mantle, Direct3D 12 and OpenGL Next will all allow VRAM stacking in multi card setups (I.E two 4GB cards in SLi/CF can work as 8GB not mirrored 4GB). If you can do it via two PCI-E slots and a connecting bus then you can definitely do it via a DIMM slot.

Don't count on it, even if someone could get it to work (not happened with Mantle yet) the bandwidth required over the PCI-E slots would be way too much.

 

[tres_kun]

That will not be enouth
A PCI Experess 16X gen 3.0 has a 126.032 Gbit/s copy paste of wiki
And a GTX 780 has a Higher memory bandwidth 336 GB/s copy of gpubos web
So you are 10.6 times slower on Pci Expres 4X compared to vram speeds

And do not forget Gen 3.0 has small visible gain compared to 2.0

 

[r7slayer]

Because GPU's tend to run out of grunt when they reach their peak Vram usage. Only in multiple GPU scenarios would this become more applicable. But there is only really a vram issue now with 4k just coming into the gaming market. But its not mature enough to actually consider anyway.

There was no real market for anyone to push this idea.

 

[Meaker]

It's actually more to do with cost. Mantle, Direct3D 12 and OpenGL Next will all allow VRAM stacking in multi card setups (I.E two 4GB cards in SLi/CF can work as 8GB not mirrored 4GB). If you can do it via two PCI-E slots and a connecting bus then you can definitely do it via a DIMM slot.

And when I say cost I mean partly the cost to implement, and also the cost of users upgrading their GPUs instead of upgrading to your newer models.

No, that's an 8Ghz bus which is connected to a chip which then routes it through the PCI express bus. The 8Ghz data rate traces are not sent straight through the PCI express bus

The main restriction on GDDR5 is the ability to maintain signal integrity at such high data rates, even now there are errors which are expected and corrected by the bus using a re-transmit.

Adding in a socket would take a dump all over the maximum memory speeds, stacked dram lets you get the ram even closer and minimise trace noise which is part of the reason it helps.

 

[Drangueos]

Well they can use physical bridges to deal with the PCI-E bandwidth limits.

 

[ubersonic]

That will not be enouth
A PCI Experess 16X gen 3.0 has a 126.032 Gbit/s copy paste of wiki
And a GTX 780 has a Higher memory bandwidth 336 GB/s copy of gpubos web
So you are 10.6 times slower on Pci Expres 4X compared to vram speeds

And do not forget Gen 3.0 has small visible gain compared to 2.0

You mean AMD lied? /shock