10GB vram enough for the 3080? Discuss..

[beany_bot]

I know nothing about cars I'm afraid

I don't think the analogy works tho. If you have a sink and only holds 3L, you can't make it hold 4L by filling it faster.

But if you make the plug hole twice the size you can fill it with more water before it overflows. It's less likely to overflow and fill said 3L

 

[willhub]

really getting fed up with the posts stating 10GB VRAM does not work this gen

 

[FoxEye]

But if you make the plug hole twice the size you can fill it with more water before it overflows. It's less likely to overflow and fill said 3L

Then you're only interested in rate of flow and overflow capacity is just there for emergencies.

Which isn't the same situation either.

 

[TNA]

really getting fed up with the posts stating 10GB VRAM does not work this gen

Haha. Maybe I should make such a thread for the ***** and giggles, but let me get my 3080 first, that will make the rage seem more genuine

 

[Poneros]

The debate is OVER! 20 GB on their way boyz

https://twitter.com/x/status/1303263760291753984

 

[Freddie1980]

Any reason to expect a 20Gb card? 16Gb would be more then enough and cheaper to produce and sell.

 

[Poneros]

Any reason to expect a 20Gb card? 16Gb would be more then enough and cheaper to produce and sell.

Yes. Memory amount is tied to bus size. So in this case for a 3080 it can only be 20 GB. There will also be 16 GB variants for 3060 & 3070 though.

 

[Grim5]

really getting fed up with the posts stating 10GB VRAM does not work this gen

The problem is not this gen LMAO

the problem is that next gen starts in 6 weeks

get with the programme and don't live in the past

 

[Demon]

When we can transfer files around using the GPU, will 10GB be enough then?

I’m just thinking about this and possibly RTX io is why 10gb will be enough..

My thoughts are that one major use of VRAM is is caching and holding textures / geometry to allow fast traversal of a scene while maintaining texture and model quality.. However the larger the world the less of it you can store in any fixed amount of VRAM, so with RTX IO and the ability to load data quickly and directly to the GPU there is less reliance on the VRAM as a caching source... Going by what the new consoles are aiming at, then I’d say 10GB is more than enough, I’d be more worried that in reality you want PCIE 4.0 to ensure max bandwidth for normal cpu/gpu traffic as well as ssd/gpu traffic which maybe make PCIE 3.0 a slight bottleneck.

 

[TNA]

The debate is OVER! 20 GB on their way boyz

https://twitter.com/x/status/1303263760291753984

Just wait? Wait for what? The price will not magically be £649 or even £699 will it.

The problem is not this gen LMAO

the problem is that next gen starts in 6 weeks

get with the programme and don't live in the past

Yea and by the time titles that start making full use of next gen come out we will be on 5nm hopper. We will soon see I guess, there are games like Resident Evil 8 which from what I understand not be on older gen consoles.

 

[otester]

With the 3080Ti it's not just about the extra 10GB but the speed as well, if I recall correctly the Ti variants were much closer to the 'titan' cards in raw speed, whether that's worth an extra $300....

 

[TNA]

With the 3080Ti it's not just about the extra 10GB but the speed as well, if I recall correctly the Ti variants were much closer to the 'titan' cards in raw speed, whether that's worth an extra $300....

Yeah, it will be between 3080 and 3090, so not much extra speed at all probably. Will depend on how much AMD push them I suppose.

 

[Freddie1980]

Yes. Memory amount is tied to bus size. So in this case for a 3080 it can only be 20 GB. There will also be 16 GB variants for 3060 & 3070 though.

The memory bus on the 3080 is 320 bit isn't it? 320 is divisible by 16 so it should work?

 

[Poneros]

I’m just thinking about this and possibly RTX io is why 10gb will be enough..

My thoughts are that one major use of VRAM is is caching and holding textures / geometry to allow fast traversal of a scene while maintaining texture and model quality.. However the larger the world the less of it you can store in any fixed amount of VRAM, so with RTX IO and the ability to load data quickly and directly to the GPU there is less reliance on the VRAM as a caching source... Going by what the new consoles are aiming at, then I’d say 10GB is more than enough, I’d be more worried that in reality you want PCIE 4.0 to ensure max bandwidth for normal cpu/gpu traffic as well as ssd/gpu traffic which maybe make PCIE 3.0 a slight bottleneck.

Even Nvidia themselves don't try to make that argument. RTX IO, which is to say, DirectStorage, is going to help with file management and improve bandwidth usage but that's it. It can not be a substitute for meagre vram amount!

SBMS-A-Man108 – Does RTX IO allow the use of SSD space as VRAM? Or am I completely misunderstanding?
[Tony Tamasi] RTX IO allows reading data from SSD’s at much higher speed than traditional methods, and allows the data to be stored and read in a compressed format by the GPU, for decompression and use by the GPU. It does not allow the SSD to replace frame buffer memory, but it allows the data from the SSD to get to the GPU, and GPU memory much faster, with much less CPU overhead.

https://twitter.com/x/status/1301134306463674370

With the 3080Ti it's not just about the extra 10GB but the speed as well, if I recall correctly the Ti variants were much closer to the 'titan' cards in raw speed, whether that's worth an extra $300....

There will be both, supposedly. A faster 3080 variant -> 3080 Ti (maybe 12GB 384bit?) AND a bigger vram 3080 variant (20GB 320bit BUT same performance)

The memory bus on the 3080 is 320 bit isn't it? 320 is divisible by 16 so it should work?

They're not gonna design a new PCB & spec when they can just re-use the already existing one. These things cost 10s of millions of dollars to do.

 

[Demon]

Even Nvidia themselves don't try to make that argument. RTX IO, which is to say, DirectStorage, is going to help with file management and improve bandwidth usage but that's it. It can not be a substitute for meagre vram amount!

Being honest nothing you’ve mentioned contradicts the idea, cache management is a combination of depth and bandwidth, the higher the bandwidth the less depth you need, and high bandwidth also hugely increases the initial seeding of the cache, something that is a disadvantage of having more depth, longer seeding times..

Imam happy to be corrected but as a SW Engineer, cache management and high speed data are something I’ve got a lot of experience with, albeit not in terms of graphics but rather real-time data analysis on embedded platforms.

 

[Poneros]

Being honest nothing you’ve mentioned contradicts the idea, cache management is a combination of depth and bandwidth, the higher the bandwidth the less depth you need, and high bandwidth also hugely increases the initial seeding of the cache, something that is a disadvantage of having more depth, longer seeding times..

Imam happy to be corrected but as a SW Engineer, cache management and high speed data are something I’ve got a lot of experience with, albeit not in terms of graphics but rather real-time data analysis on embedded platforms.

For a frame you have a limit on what the budget of your memory requirements is (in order to fully render the scene and present all the objects in that scene with their best LOD & texture). That is something that cannot be alleviated by more bandwidth. If you have less than that amount then what will happen is various objects in the scene will revert to lower LODs (at best). Bandwidth can help with the transition between LODs but if you need to hold 10 GB and you have 8 GB, then it doesn't matter how fast you can swap assets in & out because you still need 10 GB (or have to present a sub-standard image). Now if a significant portion of that 10 GB is just caching, then you'd be right, but that's not the scenario I'm talking about & not the one people are worried will happen.

 

[UEX]

The Dodge Viper would like a word
(sports car with a V10 truck engine)

In effect as opposed to contradicting my point you actually just give an example of what I'm saying. The engine is only the same in the two vehicles on paper. The cylinder size is the same. The viper v10 is not the same engine as the v10 in a truck. Same block yes, but thats it.

What I'm saying and what others, sadly few others, have pointed out is that gddr6x isn't comparable like for like with gddr6. The purpose of vram is at its core to make up for slow IO access. Now ignoring the new fancy pants future IO which will very likely reduce vram needs, the extra speed of 6x (reportedly a doubling of operations per second) over 6 should allow higher data turnover between disk and vram. Also not forgetting that system ram plays a part in this too.

So you've got the drive loading data to system ram and then on to vram the pipelines for which are not saturated at present. Id infer that 10gb is more than sufficient given the typical path of daya and that they were able to drop the total vram very easily directly because of its performance.

I think a lot of people have a very simplistic view of how vram works, like level is copied from disk and sits in vram with assets etc while you play the level. Then at the end its all cleared and the next level is loaded in. In that visualisation its insane to think that you wouldn't just need more and more vram. But its actually a very oversimplified view of a very complex process.

Like when people used to shock dead bodies to, you know, get em going again.

 

[Chuk_Chuk]

What I'm saying and what others, sadly few others, have pointed out is that gddr6x isn't comparable like for like with gddr6. The purpose of vram is at its core to make up for slow IO access. Now ignoring the new fancy pants future IO which will very likely reduce vram needs, the extra speed of 6x (reportedly a doubling of operations per second) over 6 should allow higher data turnover between disk and vram. Also not forgetting that system ram plays a part in this too.
.

Isn't the bottleneck PCIE speeds not VRAM speed?

 

[3t3P]

In effect as opposed to contradicting my point you actually just give an example of what I'm saying. The engine is only the same in the two vehicles on paper. The cylinder size is the same. The viper v10 is not the same engine as the v10 in a truck. Same block yes, but thats it.

What I'm saying and what others, sadly few others, have pointed out is that gddr6x isn't comparable like for like with gddr6. The purpose of vram is at its core to make up for slow IO access. Now ignoring the new fancy pants future IO which will very likely reduce vram needs, the extra speed of 6x (reportedly a doubling of operations per second) over 6 should allow higher data turnover between disk and vram. Also not forgetting that system ram plays a part in this too.

So you've got the drive loading data to system ram and then on to vram the pipelines for which are not saturated at present. Id infer that 10gb is more than sufficient given the typical path of daya and that they were able to drop the total vram very easily directly because of its performance.

I think a lot of people have a very simplistic view of how vram works, like level is copied from disk and sits in vram with assets etc while you play the level. Then at the end its all cleared and the next level is loaded in. In that visualisation its insane to think that you wouldn't just need more and more vram. But its actually a very oversimplified view of a very complex process.

Like when people used to shock dead bodies to, you know, get em going again.

Erm this last part is still recommended in certain instances of DEATH. I liked reading the rest though, thanks.

 

[UEX]

Isn't the bottleneck PCIE speeds not VRAM speed?

Not on pcie3 + at 16GB a second. Theres a few benchmarks out there looking at performance difference between 8x and 16x and it seems to be all but invisible and thats halving the xfer rate. I may be wrong as I'm intuiting this part without really thinking it thru but until storage access is on par with pcie capacity then the pcie bus cannot be the bottleneck really at these ram sizes.

Erm this last part is still recommended in certain instances of DEATH. I liked reading the rest though, thanks.

Ok ok....I meant long dead bodies