Will higher bandwidth of HBM memory lower VRAM usage?

So basically what I'm curious about is this: let's say a game at any particular time is using say, 4.5GB of GDDR5 memory, and let's presume that's we're not running into the VRAM limit (for example, a 6GB card).

My question is, will the same game at the same given time use less than 4.5GB of VRAM on a HBM-based GPU due to the increased bandwidth or us that a separate thing? I know it'll be better for framerate performance, but I'm curious if it'll cut down on required VRAM in the same workload scenario compared to a GDDR5-based card.

So basically (again), if you want to play/mod games that need more than 4GB VRAM, will a 4GB HBM1 GPU cut it, or is that out of the running?
I know we don't know for sure how much VRAM AMDs new cards will launch with, but I'm curious about this topic regardless.

I thought hbm was just to move it in and out of ram faster.

A texture is a texture is a texture. It will just move that texture in/out of memory faster as I understand it. So it will not replace capacity, no.

Read: http://www.anandtech.com/show/9266/amd-hbm-deep-dive

EarthDog said:

A texture is a texture is a texture. It will just move that texture in/out of memory faster as I understand it. So it will not replace capacity, no.

Read: http://www.anandtech.com/show/9266/amd-hbm-deep-dive

Click to expand...

I think what you understand about a texture file size is accurate E_D.

GPU VRAM is used to pre-load textures, they are decompressed and applied in real-time.

GPUs are data-parallel by nature, they benefit from memory architectures that can fetch
large blocks of data at a time when they begin a task on a particular dataset.

So I would think a huge data set loading in parallel generally would not allow for needing less VRAM. If it were serial data...that might be another thing.

RGone...

EarthDog said:

A texture is a texture is a texture. It will just move that texture in/out of memory faster as I understand it. So it will not replace capacity, no.

Read: http://www.anandtech.com/show/9266/amd-hbm-deep-dive

Click to expand...

Funny, I was actually skimming that article before I started the thread. Now that I've read the whole thing, I understand a bit better now.

That's unfortunate then, from my perspective, if the New AMD HBM cards only launch with 4GB. While I'm prone to sticking with team green, I was interested in seeing what AMD had to offer for a possible upgrade spree for my rig this summer, but my current and future needs over the next year or so dictate that jumping from a 3GB to 4GB card would be a poor use of $$. I was hoping to see AMD offer a 6GB card to compete with the 980ti, but it's looking like the Fury/x will probably be 4GB. I know the 390x should have 8GB GDDR5, but if that card has a TDP of 300W and it's just a rebrand....it better be real cheap by comparison

I think HBM's biggest benefit might be giving the the GPU more time to process the image (even if the identical GPUs run at the same speed) before the frame has to be transferred (extremely quickly, up to 9 times faster than GDDR5) from memory to the monitor.

It's so early in the HBM era that you may be putting too much credence on the early AMD card prospects and announcements. Before the first wave hits the streets very long the 6 and 8 GB card partners' cards will be announced and viola, problem solved.

Robert17 said:

It's so early in the HBM era that you may be putting too much credence on the early AMD card prospects and announcements. Before the first wave hits the streets very long the 6 and 8 GB card partners' cards will be announced and viola, problem solved.

Click to expand...

Hmm now that's interesting - I hadn't considered that.

However, isn't the supposed 4GB capacity an inherit limit on the first generation of HBM design? I'd think it'd be one thing for card partners to change the VRM layout and GDDR5 on a traditional "2D" PCB, but since this is just coming out of AMD's lab, would it even be possible for card partners to just be like "well we'll add some more three-dimensional stacks here and rewire X to Z and boom, we have 8GB HBM even though AMD couldn't as of now do that in-house."

I feel that, for the moment, as a new technology, that's out-of-hand for card partners.

AnimeMania said:

I think HBM's biggest benefit might be giving the the GPU more time to process the image (even if the identical GPUs run at the same speed) before the frame has to be transferred (extremely quickly, up to 9 times faster than GDDR5) from memory to the monitor.

Click to expand...

I'm thinking more along the line that the GPU won't have to wait on information to make it's calculations, or wait for large textures to be sent before it can fetch the next batch of information to be crunched. How long does information sit in VRAM before it's used? How long before it's needed is it requested, if you can move more throughput, the GPU can spend more time crunching the number than calculating when should it send for information and just ask for it when needed, GDDR5 seems to go along the lines of I can just cram the thing full of memory and it will be there when needed, HBM seems like I can get the info when needed rather than waste space holding it. Just a thought from an ignorant person hoping AMD has something up their sleeve, be nice to seem them put some pressure on "Big Green".

Robert17 said:

It's so early in the HBM era that you may be putting too much credence on the early AMD card prospects and announcements. Before the first wave hits the streets very long the 6 and 8 GB card partners' cards will be announced and viola, problem solved.

Click to expand...

Where's the fun in that?

t1nm4n said:

...HBM seems like I can get the info when needed rather than waste space holding it.

Click to expand...

See, that's what I was initially thinking/wondering/hoping for, but it sounds like that's not how it works. Even if it was, let's just take this to the extremes for example's sake, 100x faster than GDDR5, it would still need the VRAM capacity to hold the whole thing, even if it's just for the slightest moment in time, thus a moment in time that needs more than 4GB wouldn't be overcome by the alternative of excessively higher bandwidth. That's how I think I'm seeing it at the moment, but I could be totally wrong

I guess it will really all boil down to the drivers and which games they work best with. If I had any patience I would be a doctor.

It is kind of funny, a year ago hardly anybody had 4GB of vRAM, now everybody thinks that is not enough vRAM. I believe that DirectX 12 will combine all the vRAM you have for all the cards to use. Two 4GB cards will appear to have 8GB available with DirectX 12 compatible games. I think DirectX 12 compatible games are at least 2 years away.

AnimeMania said:

It is kind of funny, a year ago hardly anybody had 4GB of vRAM, now everybody thinks that is not enough vRAM. I believe that DirectX 12 will combine all the vRAM you have for all the cards to use. Two 4GB cards will appear to have 8GB available with DirectX 12 compatible games. I think DirectX 12 compatible games are at least 2 years away.

Click to expand...

Haha I know! In an increasing number of situations, though, with AAA-titles and/or 2k/4k screens and modding, it isn't enough :/ (if you're maxing settings).

DX12 probably will, but you've got to remember that not everyone SLI/Crossfires their cards, especially when doing so creates new problems that don't exist with single card solutions. Plus, as you pointed out, DX12 games will be few and far between for at least a little bit here. I mean how many games currently are even taking full advantage of DX11? (which has been out for years)

GearingMass said:

Funny, I was actually skimming that article before I started the thread. Now that I've read the whole thing, I understand a bit better now.

That's unfortunate then, from my perspective, if the New AMD HBM cards only launch with 4GB. While I'm prone to sticking with team green, I was interested in seeing what AMD had to offer for a possible upgrade spree for my rig this summer, but my current and future needs over the next year or so dictate that jumping from a 3GB to 4GB card would be a poor use of $$. I was hoping to see AMD offer a 6GB card to compete with the 980ti, but it's looking like the Fury/x will probably be 4GB. I know the 390x should have 8GB GDDR5, but if that card has a TDP of 300W and it's just a rebrand....it better be real cheap by comparison

Click to expand...

Fury has 4GB becuase type 1 HBM is limited to it for now. However looking at 3D Mark benchmarks the card is able to go neck and neck with a 12GB Titan X. Not only that but it is 100.00 cheaper than a 980ti so your price to performance is better. Becuase of it's 4GB limit though it does have diminishing returns above 4K. That will go away though once type 2 HBM is released next year. Theoretically then you could see cards from both AMD and Nvidia of 16 - 32GB. An interesting side fact as well is AMD owns the patent on HBM so Nvidia indirectly pays AMD royalties when it rolls out HBM cards next year.

Thomas_K said:

Fury has 4GB becuase type 1 HBM is limited to it for now. However looking at 3D Mark benchmarks the card is able to go neck and neck with a 12GB Titan X. Not only that but it is 100.00 cheaper than a 980ti so your price to performance is better. Becuase of it's 4GB limit though it does have diminishing returns above 4K.

Click to expand...

I'm aware of the limitations in size of HBM1, but I'm not giving AMD a pass for that. Couple of things - it might be able to run neck-and-neck with the TitanX/980ti in non-VRAM intensive applications, but as soon as you go into high-setting/high resolution scenarios, especially through the next year or two, you're looking at easily running into a wall at 4GB of VRAM. Heck, there are games right now passing that up at 1080/1440. 4GB VRAM is casting diminishing returns at and below 4K currently if you're maxing current games. And if you're buying a tier-1 GPU for gaming, do you immediately want turn down your texture settings? (And you can forget about graphical modding). Not to mention these things are going to run hot - the reference Fury X has an AIO water cooler.

Let me say this so that it doesn't just sound like I'm bashing AMD - I applaud that they're the ones ushering in HBM technology. The only problem with them being the ones using HBM1 and Nvidia jumping straight to HBM2 next year is that some people that want a high-end GPU right now don't want 4GB of VRAM when there are 6/12 available and games that can use it. However, I think that's a small percentage of AMD's consumer base, so they're probably not sweating it too much.

Thomas_K said:

That will go away though once type 2 HBM is released next year. Theoretically then you could see cards from both AMD and Nvidia of 16 - 32GB.

Click to expand...

Now that I'm looking forward to.

How does increased VRAM affect system performance, I mean I sort of always thought that you should have roughly twice the System ram for VRAM.

With this new HBM will it really be bottlenecked by available system specs, Lets put one of those new cards in my 4350FX I mean it's bad enough I'm reigning in the 290X with a 4350, but man I'd laugh tot see how poorly they worked in a subpar build, just be funny.

When HBM2 comes out is probably when it will be time to get new cards. I might be awhile before we see more improvements. That technology would be worth pending on, HBM right now (while I wish I had the cash to drop) is interesting, I think it's too early to really see the benefits, cause if DX12 does come out king, then it won't take a 1200$ card to play the best titles, a couple of 150$HBM2 cards will be just as good. One can dream right?

There is no association with system ram and the amount of vram.

HBM isn't system limited.

I guess I never understood how the information got onto the VRAM in the first place, Yes I thought it went fro HDD to RAM until it was fetched by the VRAM. Time to go learn something new.

It may go through it... (I dont think so), but there is no association between the size of your system ram vs vRAM. When vRAM is filled it will 'page out' to system ram though. I do not believe the HDD has anything to do with it.

Now, you are supposed to have 2x system ram size for page file...but even that is horsehoeey and a waste of space on your HDD. Since I have 8GB or more, I manually set the PF to 2GB. Done. Zero problems, and I saved 14GB of HDD space.

GearingMass said:

See, that's what I was initially thinking/wondering/hoping for, but it sounds like that's not how it works. Even if it was, let's just take this to the extremes for example's sake, 100x faster than GDDR5, it would still need the VRAM capacity to hold the whole thing, even if it's just for the slightest moment in time, thus a moment in time that needs more than 4GB wouldn't be overcome by the alternative of excessively higher bandwidth. That's how I think I'm seeing it at the moment, but I could be totally wrong

Click to expand...

But even though the HBM is faster than GDDR5, if it has to get textures from system RAM that data has to go over a relatively high latency 8GiB/sec PCIe bus.