Do you think we will get better frametimes and smoother gaming with HBM ?

[flopper]

always going to be bottlenecks.
you balance them out for the range of cards you build.
currently with GDDr5 you have to cache data copy across the ram that isnt needed to do with HBM. Looking at games ram requirements you find any game developer design for the userbase which is 1-2gb atm and up to 4gb.
any developer could design a 32gb ram requirement for the game but then none could play it. its why witcher 3 optimized keeps ram usage down as its designed for 2-4gb not 12 or 32gb.

Cant view HBM the way people do with GDDr5.

 

[layte]

Probably not. Cache hides the vast majority of memory latency issues in big GPU's.

 

[africanos23]

Probably not. Cache hides the vast majority of memory latency issues in big GPU's.

True this, seems to be smoother with the TXs 12GB due to more caching occuring vs other 4GB cards I had on same settings resolutions.

 

[se_vietboi]

Yes but only slightly.

 

[ginjaninja]

dunno, but feel hbm has to deliver more than bandwidth to have a practical impact on the value proposition.

 

[Rroff]

Frame times wise it depends what the bottleneck is in building and presenting new frames - often this can be CPU/RAM<>GPU and/or GPU<>GPU (multi GPU), etc. I don't think it will improve low level smoothness significantly except for in some edge cases.

And as Layte said caches, etc. often hide a lot of the real nasties memory wise.

 

[humbug]

A lot has been said about the extra bandwidth that comes with HBM, my personal view is it will have very little benefit as once you have enough bandwidth you don't need anymore.

What gets discussed a lot less is the improved latency available with HBM, this is something that I think will bring real benefits resulting in far smoother gameplay and much better frametimes.

I think when we see some frametime graphs from the new Fiji cards they are going to be very impressive indeed.

What do you guys think ?

There is a bottleneck when you consider how fast modern GPU's can render/

Smoother? yes absolutely, higher frame rates? that depends on what you are rendering.

But its not just about those things.

A GDDR5 Buffer cannot keep up, not even with a 512Bit Bus which is about as wide as its possible to get given that you need 16 32Bit IC's the make it happen.

That many IC's already has a financial cost as well as a Power-consumption cost, The memory controller needs to be big and so does the GDDR5 interface. try for a 768Bit bus and you need 24 IC's, more power, bigger IMC, bigger Interface..... and its still not enough.

Colour and texture compression helps (Proof the bottleneck exists) but its also not enough.

GPU's are becoming more and more powerful, the performance of GDDR5 is holding them back, its expensive in power consumption, die space, PCB space.

HBM is the way to go, when it arrives we will wonder why we went on for so long with GDDR5.

 

[Marine-RX179]

True this, seems to be smoother with the TXs 12GB due to more caching occuring vs other 4GB cards I had on same settings resolutions.

Funny enough, when the inevitable rebranded/reworked 290/290x comes with 8GB natively and someone claiming this being the advantage over the 4GB GTX970 and GTX980, they will be greeted by people calling it BS

It will be the typical response of things that AMD has and Nvidia doesn't have are pointless; things that Nvidia has which AMD doesn't have are benefit/advantage.

 

[humbug]

Funny enough, when the inevitable rebranded/reworked 290/290x comes with 8GB natively and someone claiming this being the advantage over the 4GB GTX970 and GTX980, they will be greeted by people calling it BS

It will be the typical response of things that AMD has and Nvidia doesn't have are pointless; things that Nvidia has which AMD doesn't have are benefit/advantage.

GTX 680 2GB - 7970 3GB
"Don't need 3GB, its pointless, 2GB is enough."

GTX 780/TI 3GB - 290X 4GB
"Don't need 4GB, 3GB is enough."

GTX 980TI 6GB - Fury X 4GB?
OMG Hyperbole only 4GB thats not enough...

 

[James J]

Doubt it'll smooth anything out unless you get a severe memory bottleneck in the current tech, which you don't unless hitting the upper VRAM limit or the 970 fiasco in terms of memory chip arrangement. HBM is like giving a wider hose pipe to the same amount of water passing through it. Unless the core can take direct advantage of the large difference in bandwidth and lower latency, it won't make a difference to performance compared to GDDR5 based tech.

 

[Jokester]

Can't say less RAM but more bandwidth fills me with any anticipation to be honest. Never really seen that much gains in overclocking GPU RAM.

 

[Rroff]

GPU's are becoming more and more powerful, the performance of GDDR5 is holding them back, its expensive in power consumption, die space, PCB space.

GDDR5 isn't really holding performance back even with the Titan X - its definitely coming to the end of the road no disputing that - this will be the last generation of GPUs where its not a factor - but AMD have jumped the gun on this one unless they have something considerably faster than the Titan X in the bag.

Can't say less RAM but more bandwidth fills me with any anticipation to be honest. Never really seen that much gains in overclocking GPU RAM.

It could be argued higher bandwidth would unlock the potential for architecture changes that would net more performance but yeah as things stand you can often underclock the memory quite a bit before performance falls off and need quite a boost on the core before overclocking memory results in any significant gains (outside of heaven benchmark).

 

[mmj_uk]

GPU's are becoming more and more powerful, the performance of GDDR5 is holding them back

its expensive in power consumption, die space, PCB space.

HBM is the way to go, when it arrives we will wonder why we went on for so long with GDDR5.

I think it's fair to say that AMD need the move to HBM much more than NVidia at this point in time, NVidia have done sufficient R&D in other areas that most of the above aren't a problem for them for the forseeable future.

I mean look at GTX980 compared to 290X, it has a memory bus only half the size but you wouldn't know it looking at the majority of benchmarks, the same will probably be the case with 980 Ti versus the 512bit 390X.

AMD's problems stem from the fact that they NEED a complex and expensive 512bit GDDR5 memory bus to be able to be competitive, they NEED whacking great VRM's for their power hungry cards, they have now reached the point where they NEED monstrous AIO coolers to sufficiently cool their GPU's whilst expelling all of the heat produced from the case. HBM is much more crucial to them going forward than it is NVidia.

 

[Sin_Chase]

I don't expect HBM to have any tangible impact on my gaming performance as a sole player. If you took GPU X with GDDR5 now and put HBM on it instead will you notice anything between the two? Doubt it.

 

[humbug]

I think it's fair to say that AMD need the move to HBM much more than NVidia at this point in time, NVidia have done sufficient R&D in other areas that most of the above aren't a problem for them for the forseeable future.

I mean look at GTX980 compared to 290X, it has a memory bus only half the size but you wouldn't know it looking at the majority of benchmarks, the same will probably be the case with 980 Ti versus the 512bit 390X.

AMD's problems stem from the fact that they NEED a complex and expensive 512bit GDDR5 memory bus to be able to be competitive, they NEED whacking great VRM's for their power hungry cards, they have now reached the point where they NEED monstrous AIO coolers to sufficiently cool their GPU's whilst expelling all of the heat produced from the case. HBM is much more crucial to them going forward than it is NVidia.

AMD chose to use a wider bus because it gives them good performance when things get stressful, its why AMD GPU's don't fall off as much as Nvidia 4K vs 1080P.

AMD also have the same Colour and Texture compression Maxwell have in Tonga, the new GPU will no doubt use it.

 

[shankly1985]

I don't expect HBM to have any tangible impact on my gaming performance as a sole player. If you took GPU X with GDDR5 now and put HBM on it instead will you notice anything between the two? Doubt it.

But then it begs the question why change from GDDR5 ?? Why would AMD and soon Nvidia make the jump if they wasn't any gains?

I just don't understand why they would do this.

 

[Sin_Chase]

But then it begs the question why change from GDDR5 ?? Why would AMD and soon Nvidia make the jump if they wasn't any gains?

I just don't understand why they would do this.

Headroom for faster GPU grunt. Supporting hardware does not necessarily have to provide a tangible benefit on it's own (IE - HBM) but it will certainly help when the other core bits of hardware (which do provide tangible benefits) improve.


I guess you can think of it like uprating your fuel pump in a performance car. On a stock engine a pump with better throughput will provide no benefit to the engine - It will feed it all the same. But when you uprate the engine and it needs more fuel, the better fuel pump will be able to provide it. In this analogy your fuel pump is memory and the engine is the GPU Core.

 

[TheMorningStar]

Is the HBM actually part of the GPU packaging? That's the impression I got. That should provide some latency gains alone.

What would be interesting is if AMD is going to be doing something like this with the CPU/APU and system RAM. I'm envisioning a transition to mega chips that handle everything that the CPU/GPU/RAM currently do as well as most if not all functionality of NB and SB also sitting somewhere in the packaging or perhaps one day in the silicone. Things seem to be getting less modular, seems to be the natural progression of improved production technologies.

 

[Mercutio]

GTX 680 2GB - 7970 3GB
"Don't need 3GB, its pointless, 2GB is enough."

GTX 780/TI 3GB - 290X 4GB
"Don't need 4GB, 3GB is enough."

GTX 980TI 6GB - Fury X 4GB?
OMG Hyperbole only 4GB thats not enough...

Fairly quote-worthy to be honest. That's exactly how it's going to go.

 

[mmj_uk]

It doesn't factor in time though, 12mths is an eternity in computing terms.

2GB was fine when GTX680 released.
3GB was fine when GTX780Ti released.

4GB will probably be fine when Fury releases but it won't be long before games are pushing beyond it, considering the whole idea of HBM is to take memory performance to the next level (primarily for the extreme/4K users) it's pretty stupid that it'll be borderline not enough on release when 8GB of GDDR5 would have probably been more optimal in the long term at least until such a time that 8GB HBM was possible.