GPUs with Expandable VRAM!

[Vathrex]

This existed before. I read it was common in the early 90s. There would be slots for extra vram chips. The S3 Trio had it

[exclar]

The main reason they haven't done this is that it would make the RAM slower. As you can see, the RAM chips on the GPU are very close to the GPU chip itself to reduce latency. This is why this is not recommended.

[GhostyKingdom]

Engineer here. Unfortunatly, this cant happen (yet) because ddr6 VRAM is highly specialized and not commercially available and tends to be built differently between GPU's.

[Tempestan]

As others have posted, yes, they had this feature back in the late 90s and early 2000s. There is a reason why you do not see them on cards today.

[wesleyhalliwell]

They have done this, in fact it used to be quite common. The reason it stopped was the same reason why it’s difficult to find the highest speeds of DDR5 in laptops with slotted memory: it was becoming too difficult to increase the speed of the memory with the much longer traces that you need for slots. When the memory is soldered directly to the board, the traces can be much shorter because they’re able to put the chips right next to the GPU. That’s why you see a circle of memory chips around the GPU if you take off the heatsink of a modern graphics card, it allows them to all be as close as possible and the traces as short as possible.

[dant]

Expendable VRAM was a thing on workstation cards a long time ago. Welcome to Tech, Zach.

[DigBipper]

There are 3 core reasons why socketed / expandable VRAM and GPU dies aren't a thing.

1: Cost

It's more expensive to add a socket that the chips seat into than creating a BGA footprint and soldering the parts directly onto the board. It also takes more manufacturing time and adds complexity. Both drive up the end cost.

2: Latency

The further away an electrical signal must travel, the longer it takes for it to make it from point A to B. In RAM this causes latency. This increased latency reduces the end performance of the memory and therefore can slow the GPU if it is performing latency sensitive tasks (which it's always doing)

3: Reliability

Increased trace length, plus the existence of the socketed interface both introduce the potential for the introduction of noise on a signal. This noise has the potential to cause memory errors and therefore would increase the chances of corruption or general hardware instability. GDDR memory, especially newer standards such as GDDR5 and GDDR6 are both extremely high speed, high bandwidth components. This high speed and bandwidth makes them extremely susceptible to noise. It's better to keep the parts right next to the GPU package soldered directly to the board for these reasons. Soldering creates a stronger electrical connection than socketing and since you don't have added distance it reduces the chances of an interference induced error.

[triynizzles]

They would have to completely redo, the memory controller architecture. Currently if you have a 64 bit bus(small number to make the example easier), then there’s two chips on the pcb. The only thing you could do is potentially clamshell to double size, but maintain speed or go from 18gbps to 19gbps (faster frequency) but the main thing that will give you more bandwidth and vram size is a larger bit bus, so there are more chips providing data simultaneously.

[bush]

They have done that, back in the days some gpu's had sodimm-like slots. That was discontinued due to the distance from the memory chips to the dye. Every additional milimiter can cause performance drops

[ossiaaltola]

The reason that they don't have expandable VRAM is that it would make the traces from the VRAM to the gpu die too long. The closer the VRAM is to the GPU the faster you can make the ram. And also they want your money.

[hippityhop]

While normal RAM isn't really affected by distance from CPU, VRAM is very dependant on the distance between it and the core of GPU.

[icerink]

Being detachable increases physical dimensions of the connection which also increases the latency of the memory

[jmtradbr]

They had in the past, the technical reason they don't have today is speed. The closer the ram is to the die, the better. Also all the vram chips needs to have the trails at same length. Adding removable vram will affect the solded vram.

[notisac]

That would be amazing, especially as newer games require sooo much more vram. There would also be literal tons less e-waste as older cards would remain viable for possibly years longer!

Edit: It seems this used to be a thing on older cards. The prevailing thought is that expanded vram is quite a bit slower. Not sure how much validity there is in that but I’m sure there’s at least some truth to it. Still, it’s fun to see people replace parts in older cards to install more vram themselves, just not as quick and easy as a little slot like an sd card or what have you.

[LexyDaShmexxy]

they used to do that but there was a limitation in the speed and bandwidth, it's just too slow to implement, that's why vram chips are so close to the gpu die

[kristofmadarasz]

Fun fact :Back in day's they did expandable vram on gpu's just they stoped using it because it was slow and expensive .

[BF]

Expandable memory GPUs actually existed a while ago... the S3 Virge had 2MB on board and you could add a daughter board with another 2MB of memory :) I still have it in my DOS Stuff box 😇

[SkylandPirate]

Hey Zach, comp engineering student here. Watching your channel is giving me ideas on what I could one day do. Imma just take ideas and hopefully one day make it.

[verhulstak]

They actually used to do that but a socket cant do the speeds for it

[PursuerOfTruth]

No... I do wanna see ram slots sticking out of the top. It'd be like those fins on a race car and I think with the right design it could be sick.