Extra VRAM GPUs - Factual / In-Depth Discussion

[Bonn]

So I've recently seen an article in the news section for a "BGF Titan X" with 24GB of VRAM. As I some previous forums posts around overclock.net went, the marketing around these types of cards is cruel and sly for the average users and consumers.

 

Let's throw a discussion out out the about these double memory cards. 

 

Firstly, some sources on memory controllers; http://international.download.nvidia.com/geforce-com/international/pdfs/GeForce-GTX-750-Ti-Whitepaper.pdf

 

http://www.tomshardware.com/reviews/nvidia-geforce-gtx-960,4038.html

 

The GTX 750 Ti could be a good example, however OEMS would lose out doubling memory on these. There's more information on Fermi type cards, however close to irrelevant now. I'll use the GTX 960 as an example here as the 4GB cards seem to be streaming out now.

 

A GTX 960 has the following on die. Two graphics processor clusters. Each cluster contains 1x 64 Bit memory controller ( two in total for the card ) and 4 SMMS ( 8 SMMS of 128 Cuda cores ). That's all we care about here.

 

Now, Nvidia create the card with their memory config which is a viable option for their own reasons, price to performance and those sort of things. For the GTX 960 the reference is 2GB of VRAM. You then divide this up into your physical memory chips found around the GPU on the card and you get 8x 256mb GDDR5 memory chips that then pipe though into the two 64 bit memory controllers on-die. Performance of the GPU core is designed around this and specifically around the memory controllers. 

The problem we have is that OEMs then release their own "improved cards" which they stuff around with the PCB and bits they're allowed to play with. This means memory. An easy way to make cheap money and gain customers is to sell something the others don't have. So, after a while we see these GTX 960's with 4GB of VRAM! How do they do it!?

Nobody ever looks or asks why, and it seems to be palmed off in the corner with lots of well known tech pages. The OEMS swap or double the memory chips. Either way, this creates a bottleneck.

Explaining below this is what happens.

A stock 2GB Card.
You have two lanes ( 2x memory controllers ) of highway into the car park. Two cars can enter the car park at the same time efficiently.

 

A OEM 4GB Card.
You have two lanes ( 2x memory controllers ) of highway into the car park, but now you have 4 cars wanting to enter the car park, this creates a bottleneck, but it's usable, however efficiency is out the window, and latency is brought into the equation. Latency + games + vram = bad.
 

OEMS can't modify the GPU, memory controllers can't be screwed with or swapped, they live in the GPU die, but the memory lives outside which the OEM can then play with.

Until we get stacked VRAM or AMDs joke of HBM, this problem will still be present from OEMs.

Running the memory performance testing on a 2GB Card reference and a 4GB non-reference card will show it's usable, however these are not real-world reflective benchmarks. Many other posts around the internet have stuttering issues with OEM non-reference memory cards. This is due to the highway and car park scenario above. If you're gaming which most of us are using GTX cards for, this is only wasting money, for sometimes worse performance.

I want this post to turn into a discussion, and if people have these cards around, then lets get testing this and get results and share it. My aim here is to save everyone here, or somebody that's researching a build or what to buy next a penny or seven.

 

[Dabombinable]

Well typed, I'll have a look at the configuration of my 2GB GTX650ti before saying anything else.

[Johannes_Lazor]

A stock 2GB Card.

You have two lanes ( 2x memory controllers ) of highway into the car park. Two cars can enter the car park at the same time efficiently.

 

A OEM 4GB Card.

You have two lanes ( 2x memory controllers ) of highway into the car park, but now you have 4 cars wanting to enter the car park, this creates a bottleneck, but it's usable, however efficiency is out the window, and latency is brought into the equation. Latency + games + vram = bad.

 

Just becouse a card has twice the Vram doesnt mean that it Has to store the data twice as fast if i understand it correctly.

It is rather the size of the parking lot that changes, Not the number of cars that drive into it at the same time.

[Johannes_Lazor]

In my opinion Nvidia should really have put 3gb of Vram in the 960 according to me since Next year it will probably have hard competition against the next gen R9 370x(Our current R9 280x) since games will most likely require more Vram.

Edited March 21, 2015 by Johannes_Lazor

[Dabombinable]

-snip-   

[Bonn]

It's more about the latency or the time to get your car from the lot. This where you get stutter in things like games.