Computerbase - 3gb vs 4gb vs 6gb vs 8gb GDDR5 VRAM Frametime Testing

[Phynaz]

Huh? It's paging 4GB to system ram. The Titan is 30% faster.

 

[Bacon1]

Huh? It's paging 4GB to system ram. The Titan is 30% faster.

30% faster... wow thats one amazing way to skew the FPS #s lol

Neither are playable at 13 vs 10 fps. The whole point is MEMORY isn't limiting even at 4k 4x SSAA.

 

[96Firebird]

Why do you think that has to do with the memory being HBM rather than AMD focusing on memory management?

 

[Thinker_145]

And yet the fury can't play Doom at nightmare settings. This is all a pointless debate. The fury is one of the worst high end cards AMD has launched in a long time, just move on.

 

[Bacon1]

Why do you think that has to do with the memory being HBM rather than AMD focusing on memory management?

Because the huge amount of bandwidth from HBM allows for that memory management... How well does the 980 TI hold up @ 4k 4x SSAA in ROTTR?

 

[Bacon1]

And yet the fury can't play Doom at nightmare settings. This is all a pointless debate. The fury is one of the worst high end cards AMD has launched in a long time, just move on.

Steam Launch Option "+menu_advanced_AllowAllSettings 1" unlocks Nightmare mode

 

[96Firebird]

Because the huge amount of bandwidth from HBM allows for that memory management...

How so?

 

[Headfoot]

What is publicly available knowledge IIRC is that AMD is doing special memory management for Fury. Does it rely on HBM's bandwidth to be feasible? That is not something I have seen any published information on. I would imagine they are using generalized memory management techniques like caching / paged memory which doesn't rely on HBM's bandwidth advantage except to make it a little faster.

 

[Phynaz]

Because the huge amount of bandwidth from HBM allows for that memory management... How well does the 980 TI hold up @ 4k 4x SSAA in ROTTR?

HBM has nothing to do with using system memory to hold graphics data. Accessing system memory for textures has been a thing since AGP was introduced in 1997.

 

[UaVaj]

these results are nothing more than useless benchmarks. any frame time above 17ms is diminished real world gaming enjoyment.

reduce game setting to maintain frame time 17ms or less and 2gb vram will be sufficient (at least for 1080p).

 

[Thinker_145]

Steam Launch Option "+menu_advanced_AllowAllSettings 1" unlocks Nightmare mode

And any performance benchmarks for that?

Sent from my HTC One M9

 

[kraatus77]

just look at l2/1/3 cache, slower = larger size , faster = smaller size.

that's why faster HBM can make up for smaller vram. but guess it's too difficult to grasp because it's the amd who is doing it ? and you green fans feel ashamed for some stupid reason ? lol


Trolling is not allowed
Markfw900
Anandtech Moderator

 

[Dribble]

HBM is just memory like DDR, nothing more, nothing less. It's not magically special and doesn't allow you to have more memory then DDR. The reason fury works a bit better then 4GB would suggest is AMD spend a lot of time trying to work around the 4GB limit in drivers or no one would buy it.

 

[96Firebird]

just look at l2/1/3 cache, slower = larger size , faster = smaller size.

that's why faster HBM can make up for smaller vram. but guess it's too difficult to grasp because it's the amd who is doing it ? and you green fans feel ashamed for some stupid reason ? lol

This whole post is completely wrong.

 

[Mondozei]

This whole post is completely wrong.

You're free to spell out why instead of providing a fact-free oneliner.

 

[Thinker_145]

Why don't someone benchmark a 980 and Fury in Doom nightmare? If only the Fury manages to do well then we know it's HBM. If neither cuts it then we know it isn't. If both work fine then we just know it's an artificial limitation in the game.

 

[96Firebird]

You're free to spell out why instead of providing a fact-free oneliner.

It's pointless trying to explain for the umpteenth time. If someone does it get it by now, they aren't going to get it.

 

[maddie]

What Bacon1 and kraatus77 are saying is that if a GPU needs a certain bandwidth to function at full speed and which is less than the available bandwidth, you can use the excess bandwidth to allow for VGA memory access to main memory without interrupting the GPU > VGA memory transfer.

This should bypass the traditional effect of interrupting to GPU >VGA memory access to allow VGA memory > main memory access, which normally causes stuttering.

It appears possible with the use of caches to buffer transfers.

Is this happening? I don't know, but it appears possible, and will be more apparent at higher resolutions.

 

[Carfax83]

just look at l2/1/3 cache, slower = larger size , faster = smaller size.

that's why faster HBM can make up for smaller vram. but guess it's too difficult to grasp because it's the amd who is doing it ? and you green fans feel ashamed for some stupid reason ? lol

This is like the worst analogy ever! Hint, textures don't care about speed or bandwidth. They care about capacity. Doesn't matter how fast the VRAM is, if the textures routinely exceed the capacity, then the GPU will have to swap to system RAM, which incurs a massive latency penalty..

Larger VRAM pool means less texture swapping, which means smoother frame delivery. It ain't rocket science..

And any performance benchmarks for that?

Sent from my HTC One M9

This was done at 4K, so AMD's poor CPU overhead in OpenGL shouldn't really matter. Just look at the RAM usage between the two, it's obvious the R9 Fury X is having to swap textures to system RAM way more often than the GTX 980 Ti. The Fury X is unplayable, but the GTX 980 Ti is still playable.

Looks like the Fury X wasn't even using it's full 4GB framebuffer though.

 

[96Firebird]

The bottleneck (which causes the stuttering) is transferring over the PCI-E lanes and to system RAM/SSD/HDD. It doesn't matter how fast the VRAM can go, that isn't where the bottleneck is.

 

[Carfax83]

What Bacon1 and kraatus77 are saying is that if a GPU needs a certain bandwidth to function at full speed and which is less than the available bandwidth, you can use the excess bandwidth to allow for VGA memory access to main memory without interrupting the GPU > VGA memory transfer.

This is completely wrong. GPU to system RAM transfers are limited by the PCIe bus, which allows 16GB/s of bidirectional transfer with PCIe 3.0. Doesn't matter how fast the VRAM is, it's going to run into this bottleneck once it pages to system RAM. Even worse is the latency however.

This is why VRAM capacity matters, because the more VRAM you have, the less trips the GPU has to make to system memory. A well tuned 3D engine and the OS will balance texture swapping between VRAM, RAM and storage so that the latency penalty is hidden.. But if the VRAM limit is constantly being exceeded, then there is nothing that can rectify it other than getting more VRAM.

 

[maddie]

Is it possible to do a burst transfer from VGA memory to an on-die cache that reduces the needed memory access time when compared to a slower memory?

The free time now available should allow for a main memory access for swapping data without slowing the GPU.

This should work with a bandwidth in excess of that normally needed.

Some seem to think that you must do the data transfer in one operation [which will cause a noticeable to humans delay AKA stuttering] instead of splitting it into micro units which should be transparent to us.

Is this possible or impossible?

 

[Carfax83]

Is it possible to do a burst transfer from VGA memory to an on-die cache that reduces the needed memory access time when compared to a slower memory?

GPUs have caches and they also have tons of registers. However, for this particular instance, those would be useless as textures take up way too much space to be stored in cache or registers. Cache and registers are used for much smaller sized data.

 

[maddie]

This is completely wrong. GPU to system RAM transfers are limited by the PCIe bus, which allows 16GB/s of bidirectional transfer with PCIe 3.0. Doesn't matter how fast the VRAM is, it's going to run into this bottleneck once it pages to system RAM. Even worse is the latency however.

This is why VRAM capacity matters, because the more VRAM you have, the less trips the GPU has to make to system memory. A well tuned 3D engine and the OS will balance texture swapping between VRAM, RAM and storage so that the latency penalty is hidden.. But if the VRAM limit is constantly being exceeded, then there is nothing that can rectify it other than getting more VRAM.

Yes, this is true, but what I'm saying is that a very high bandwidth memory moves the point at which this happens. You can, to some degree make allowances for memory size by having excess bandwidth. Eventually you will saturate the new mechanisms, but this will always happen for all cards, no matter the memory.

I don't know, but are there any cards that will buffer the entire [complex] game in one go?

 

[Dribble]

Yes, this is true, but what I'm saying is that a very high bandwidth memory moves the point at which this happens. You can, to some degree make allowances for memory size by having excess bandwidth. Eventually you will saturate the new mechanisms, but this will always happen for all cards, no matter the memory.

No it doesn't, HBM does not make the PCI-E bus magically faster. 4GB of VRAM is 4GB of VRAM, if it fits it fits, if it doesn't it goes to system memory via the PCI-E bus. That PCI-E bus runs at the same speed irrespective of your video card and what type of memory it has.