Why VRAM will never stack in SLI - regardless of DX12

[FruitBasketSilex]

Okay, but what if the access to the memory was done through a proprietary connector, I'm thinking of a wider, much faster SLI bridge acting as the the memory bus.

I'm thinking of Mountain Dew branded GPUs.

[Guest]

Those jokes are old as well.

Yep. The Nvidia memory joke is relatively new. So I don't even.... Just ignore it something.

Because look at it this way, people are socializing and then somebody makes a mainstream albeit overused joke and suddenly somebody is like 'don't even start' and is serious about it.

That makes me think there is something going on.

[FruitBasketSilex]

Okay, but what if the access to the memory was done through a proprietary connector, I'm thinking of a wider, much faster SLI bridge acting as the the memory bus.

I'm thinking of Mountain Dew branded GPUs.

[Guest]

I've seen plenty of people that like AMD GPUs. I've seen plenty of people reccomend AMD GPUs instead of Nvidia ones.

However, nvidia does have some pretty nice shit and quite recent, with low power draw, and AMDs big thing right now is "Hey, throw money at us while you wait for us to finally release a new product" (in reference to the (possibly) new FX chips, zen, and 300 series GPUs)

It isn't so much about the Nvidia vs amd thing.

I thought he was a fanboy or something that couldn't take a joke about his company.

But it seems he isn't one.

Also you're also being an ass for making me out to be an ass because I see something unusual and point it out in a rough way.

[FruitBasketSilex]

Yep. The Nvidia memory joke is relatively new. So I don't even.... Just ignore it something.

Because look at it this way, people are socializing and then somebody makes a mainstream albeit overused joke and suddenly somebody is like 'don't even start' and is serious about it.

That makes me think there is something going on.

I believe a large part of that is due to the lack of an ability to use inflections and tones. If I were to say to you "I'll fuck your mother" IRL, I could make it sound very much like a joke. If I say I'll fuck your mother on here, then you're already on eBay looking for cheap chastity belts.

Point: large amount of it is due to only being able to communicate through text.

[Godlygamer23]

Yep. The Nvidia memory joke is relatively new. So I don't even.... Just ignore it something.

Because look at it this way, people are socializing and then somebody makes a mainstream albeit overused joke and suddenly somebody is like 'don't even start' and is serious about it.

That makes me think there is something going on.

I've been complaining about the 3.5GB VRAM thing for a while.

[Guest]

I believe a large part of that is due to the lack of an ability to use inflections and tones. If I were to say to you "I'll fuck your mother" IRL, I could make it sound very much like a joke. If I say I'll fuck your mother on here, then you're already on eBay looking for cheap chastity belts.

Point: large amount of it is due to only being able to communicate through text.

That is also true...

[FruitBasketSilex]

It isn't so much about the Nvidia vs amd thing.

I thought he was a fanboy or something that couldn't take a joke about his company.

But it seems he isn't one.

Also you're also being an ass for making me out to be an ass because I see something unusual and point it out in a rough way.

I did that because rough on here usually translates straight to flamewars.

[Guest]

I've been complaining about the 3.5GB VRAM thing for a while.

Yeah okay, you don't even have to complain about it, it's not what I am saying. But just thought you were something else.

But okay that's settled.

[Godlygamer23]

It isn't so much about the Nvidia vs amd thing.

I thought he was a fanboy or something that couldn't take a joke about his company.

But it seems he isn't one.

Also you're also being an ass for making me out to be an ass because I see something unusual and point it out in a rough way.

I am not a fanboy of any company. NVIDIA deserves criticism for what they did to its customers with the 970. I also happen to have an AMD APU running my server.

[Guest]

I am not a fanboy of any company. NVIDIA deserves criticism for what they did to its customers with the 970. I also happen to have an AMD APU running my server.

Yes but really it isn't about what you use for me. Really doesn't matter to me.

Simply thought you were an overzealous fanboy and that's it. I just used AMD heat jokes argument as a comparison to 'lure' you out should you really be one.

[FruitBasketSilex]

Yes but really it isn't what you use for me. Really doesn't matter to me.

Simply thought you were an overzealous fanboy and that's it. I just used AMD heat jokes argument as a comparison to 'lure' you out should you really be one.

Such tactics much amaze

[Guest]

Such tactics much amaze

Hey it works and it isn't a dirty tactic so.... Fair game.

[FruitBasketSilex]

Hey it works and it isn't a dirty tactic so.... Fair game.

I just wanted to make a doge reference.

[Guest]

I just wanted to make a doge reference.

I know

[Emanuel_M]

vram does not stack or combine together. It does not work like that =/

[FruitBasketSilex]

I know

<3

[FloRolf]

So, pcie speeds are doubling each gen.

That means we need pcie 7.0 for 240gb/s

But at the time we have 7.0 gpus all have hbm technology which (now) is 5x faster than ddr5.

That makes memory run at over 1gb/s so we need pcie 9.0 or 10.0

Gg ez. #fml

[Sam Z Man]

Okay, but what if the access to the memory was done through a proprietary connector, I'm thinking of a wider, much faster SLI bridge acting as the the memory bus.

Me and my friend were talking about this last week we saw this coming because we thought how is GDDR5 to be able to though a pci slot at any reasonable speed at 16x let alone 8x we then determined that we should talk to the chief AMD GPU developer turns out our physics teacher was a good friend of his in university and sujest that they should develop GPU's that have 6 inch wide bridges that have hundreds of pins to enable vram stacking. We already gave up hope on his dual 970s or my dual 290Xs ever getting 8gb. I always viewed DX12 vram stacking with the attitude of I will believe it when I see it.

[Cheddle]

Okay, but what if the access to the memory was done through a proprietary connector, I'm thinking of a wider, much faster SLI bridge acting as the the memory bus. 

 

Take a look at the RAM on a graphics card, see how many individual pins there are? these are not just for show... each pin can only carry so much data, combining all of these pins into some kind of enormous parallel cable is just technically VERY difficult to do, for the same reason that we moved form ParallelATA ribbon cables to SerialATA cables for harddrives, went from 48-pins to 7-pins and got better speeds... 

 

http://www.computerworld.com/article/2574104/data-center/serial-vs--parallel-storage.html

 

The above difficulties apply when trying to design a proprietary connector. Nvidia has something in the pipeline to alleviate this, being NVlink, but its not likley to reach consumer land for some time

 

----

 

To all the 3.5gb clowns, imagine all "8" Gbs of your SLI'ed 4gb cards being even more gimped than that last 512mb of RAM on the 970... thats how sharing RAM over PCIe is going to go for you.

[guyf2010]

I'm not sure if the OP is serious or not but regardless I will go on.

 

If all the current data that goes to a single card is fast enough, but is only connected through PCIE, where is this extra bandwidth coming from?

 

Each card has it's own PCIE bandwidth, thus having two independent cards should not cause any issue. As far as I am aware, the idea of the way DX12 works is each independent card can be accessed and sent individual instructions without communication with another card, just the CPU. DX12 pools available hardware and manages which parts of the screen are rendered by each card.

 

Also, to my knowledge, the 224GB/s or so that is the "maximum memory bandwidth" is bandwidth between the Graphics Processing Core and the memory. I could be wrong, but there is no bus that can load data into the memory for a single card at > 50GB/s.

 

TL;DR: Why would PCIE bandwidth be a massive issue for Multi GPU setups and not single cards? Data is loaded into a single GPU's memory through PCIE as well.

[Sauron]

A hard drive is already a bottleneck to me, and all I do is limited multitasking around windows and booting (already into OS, waiting for programs and services to load)

 

I meant specifically bottlenecks in games with texture loading and offloading, which for now aren't an issue.

[Sauron]

I'm not sure if the OP is serious or not but regardless I will go on.

 

If all the current data that goes to a single card is fast enough, but is only connected through PCIE, where is this extra bandwidth coming from?

 

Each card has it's own PCIE bandwidth, thus having two independent cards should not cause any issue. As far as I am aware, the idea of the way DX12 works is each independent card can be accessed and sent individual instructions without communication with another card, just the CPU. DX12 pools available hardware and manages which parts of the screen are rendered by each card.

 

Also, to my knowledge, the 224GB/s or so that is the "maximum memory bandwidth" is bandwidth between the Graphics Processing Core and the memory. I could be wrong, but there is no bus that can load data into the memory for a single card at > 50GB/s.

 

TL;DR: Why would PCIE bandwidth be a massive issue for Multi GPU setups and not single cards? Data is loaded into a single GPU's memory through PCIE as well.

 

You're right, that's exactly where the OP's supposition falls.  The 224gb/s bandwidth is exclusively between the gpu and vram, the data still streams through pcie while loading a scene, which means it won't limit rendering times (if a game is coded properly of course).

[AsGoodAsAHacker]

Same here with my 970s, 7GB 8GB for me!

7gb and 1gb snail mem

[Cheddle]

I'm not sure if the OP is serious or not but regardless I will go on.

 

If all the current data that goes to a single card is fast enough, but is only connected through PCIE, where is this extra bandwidth coming from?

 

Each card has it's own PCIE bandwidth, thus having two independent cards should not cause any issue. As far as I am aware, the idea of the way DX12 works is each independent card can be accessed and sent individual instructions without communication with another card, just the CPU. DX12 pools available hardware and manages which parts of the screen are rendered by each card.

 

Also, to my knowledge, the 224GB/s or so that is the "maximum memory bandwidth" is bandwidth between the Graphics Processing Core and the memory. I could be wrong, but there is no bus that can load data into the memory for a single card at > 50GB/s.

 

TL;DR: Why would PCIE bandwidth be a massive issue for Multi GPU setups and not single cards? Data is loaded into a single GPU's memory through PCIE as well.

 

You can't render different parts of a scene on different cards without each card having the full resources of the scene, or communicating with the other card, sure for some workloads you could but shaders don't work like that... HDR, bloom, motion blur, for example all require the entire frames data to render, Im not a programmer but as far as I know it cannot be done until the completing card has the entire frame ready.

 

Regardless of if the CPU is issuing instructions to each GPU individually, the only way to reduce the VRAM usage is to load resources into each card asymmetrically. If one cards workload suddenly requires the resources it has not yet loaded then it will be bottlenecked by the exact same thing that currently bottlenecks a card when it runs out of VRAM

 

You're right, that's exactly where the OP's supposition falls.  The 224gb/s bandwidth is exclusively between the gpu and vram, the data still streams through pcie while loading a scene, which means it won't limit rendering times (if a game is coded properly of course).

 

You are both imagining that the initial data is the same as the working data.

 

Based on your ideas you may as well not have VRAM on a card, just have it all accessed over PCIE, that will work well...