HardOCP looking into the 970 3.5GB issue?

[Flogger23m]

Ok so I wanted to do a test with Dragon Age since I played the fuck outta it (usually with 2xmsaa @ 1440p).

So I decided to put 4xMSAA @ 1440p everything maxed.....and wow yea I can notice the stuttering when getting above 3.5gb. I honestly thought this could only happen to certain/card systems, but damn I can actually reproduce it......Crazy indeed.

So the reason @ 4k this isnt a problem is because the GPU is already pegged at 99%?, but if the GPU is around 70-80% usage, and over 3.5gb I get horrible stuttering.....wow this is crazy. I was going to say it happens at certain situation, but with more and more games need more vram, this is going to be a serious issue in the near future.

What are your average frames? The game does seem to be demanding, so maybe the GPU just can't keep up with that res and AA? But if you're getting 50+ frame rates with large stutters that may point to an issue.

BTW, what brand/model is it?

 

[misterbobby]

Context? The you toss around the word allocation like it has a specific meaning... Like Dynamic or Static. Your context isn't specific enough for people to even understand what you trying to explain. The problem isn't unused allocated memory. The problem is that the game fully utilizes the allocated RAM it's been allowed to use. Stop with the semantics and understand the issue here. It's a fact, people are getting micro stuttering issues. Even at 1080p settings on a GTX 970.

Here is my source for the memory usage claim:
http://www.guru3d.com/articles_page..._graphics_performance_benchmark_review,8.html

This should not even have to be explained at this point. Its ridiculous that you dont know that games can use more vram if you have but they dont always necessarily need it. In fact in most cases they dont. Using your logic, all of my games that use over 3gb of vram on my 980 are not playable on my 780.

 

[Brackle]

What are your average frames? The game does seem to be demanding, so maybe the GPU just can't keep up with that res and AA? But if you're getting 50+ frame rates with large stutters that may point to an issue.

BTW, what brand/model is it?

In sig. I had 970 SLI when I tested, both gpu's around 70% usage.

 

[misterbobby]

In sig. I had 970 SLI when I tested, both gpu's around 70% usage.

Yeah with SLI I can certainly see this being an issue in some vram demanding games as you are able to crank more details at higher resolutions.

 

[RainStryke]

This should not even have to be explained at this point. Its ridiculous that you dont know that games can use more vram if you have but they dont always necessarily need it. In fact in most cases they dont. Using your logic, all of my games that use over 3gb of vram on my 980 are not playable on my 780.

So you are saying RAM "USED" is "Allocated" RAM? I'm starting to think you don't understand what allocation means... If .5GB of that 3GB of RAM on the GTX 780 was faulty, yes... It would be micro stuttering like what GTX 970 users are experiencing. At that point, what is playable or not is objective.

 

[Brackle]

Yeah with SLI I can certainly see this being an issue as you are able to crank more details at higher resolutions.

The reason I never noticed it before, is because I was trying to get as high as framerate as possible since I run 120hz.

So when I try to keep my frames as high as possible I usually turn off AA, so I never noticed it. Probably wouldn't of even tried it unless this issue came up.

Well Yea it really did show up for me. I love my smooth framerate, and I HATE stuttering (which is why I left AMD Crossfire before XDMA they use now).

Anyway yea...pretty shocking.

 

[misterbobby]

So you are saying RAM "USED" is "Allocated" RAM? I'm starting to think you don't understand what allocation means... If .5GB of that 3GB of RAM on the GTX 780 was faulty, yes... It would be micro stuttering like what GTX 970 users are experiencing. At that point, what is playable or not is objective.

You were saying if a game shows a certain amount of vram being used that the game actually needs it. That of course is false. I said according to your logic then a game using over 3gb of vram on a 980 would not be playable on a 780 then. Now your bizarre reply is that unless .5 gb of the 3 gb on the 780 was faulty it would be? Lol now you are making even less sense as again your earlier stance was that games were using more vram because they must need it.

 

[RainStryke]

It's not a matter of need, it's a matter of if it's used. Once a GTX 970 tries to use more than 3.5GB of RAM because it's been allocated 4GB, it will have these issues.

 

[evilsofa]

So you are saying RAM "USED" is "Allocated" RAM? I'm starting to think you don't understand what allocation means... If .5GB of that 3GB of RAM on the GTX 780 was faulty, yes... It would be micro stuttering like what GTX 970 users are experiencing. At that point, what is playable or not is objective.

Some games preload textures into VRAM before those textures are needed. Those textures may or may not actually get used during the gameplay session. I know that Battlefield 4 does this, and there are probably others now.

The trick with those games is that we don't necessarily know what order they load textures into VRAM. Perhaps some games will load the most commonly used textures in first and the most rarely used textures in last, which could result in this problem being rarely or never triggered. But others might do it in alphabetical order, or by creation date, or some other organizational criteria that is not easily discerned and cause a commonly used texture to be put on the 0.5GB VRAM section, sometimes or all the time.

 

[misterbobby]

It's not a matter of need, it's a matter of if it's used. Once a GTX 970 tries to use more than 3.5GB of RAM because it's been allocated 4GB, it will have these issues.

And I am not seeing that anywhere. In fact all the testing I have seen shows that most games go right past 3.5 with no issues at all. It seems only games that are played on settings that actually "need" over 3.5 are having the issue.

 

[RainStryke]

I'm guessing you missed the part where Nvidia responded to the problems, mentioning that they partitioned .5GB of the memory and memory tests show that the performance of the partitioned memory is slower than molasses. They've cut corners and they are being called out on it... They have not directly responded to Nai's benchmark yet. Which is currently the proof behind the performance of the allocated memory.

 

[misterbobby]

I'm guessing you missed the part where Nvidia responded to the problems, mentioning that they partitioned .5GB of the memory and memory tests show that the performance of the partitioned memory is slower than molasses. They've cut corners and they are being called out on it... They have not directly responded to Nai's benchmark yet. Which is currently the proof behind the performance of the allocated memory.

Um of course I have seen the reply and their tests showed going over 3.5 made an additional 1% to 3% performance drop.

 

[RainStryke]

A 3% performance difference between the GTX 980 and GTX 970 ratio?

The problem is, I had micro stuttering with a solid 90FPS reading on EVGA precision in game monitor... So i'm not really sure how their FPS tests are really going to reflect the majority of the GTX 970 users issues... Looking at my average frame rates, you would wonder how I was getting bad performance... But dealing with it in person is a whole different experience.

 

[Tamlin_WSGF]

This should not even have to be explained at this point. Its ridiculous that you dont know that games can use more vram if you have but they dont always necessarily need it. In fact in most cases they dont. Using your logic, all of my games that use over 3gb of vram on my 980 are not playable on my 780.

Even if a game can use more then 3GB of memory, it doesn`t mean that it cannot use less then 3GB of memory. As evilsofa point out, a game can preload textures into the available memory that the card have.

While you play the game, the textures can be swapped with other textures. I think the problem here is that 500MB of the 970 has low bandwidth, so any data stored on that area or data swapped on that area might slow things down causing framedrops and stuttering when used.

On those games, you`d probably be better off if Nvidia wouldn`t expose those 500MB to the game and instead just have 3.5GB available for the game instead.

 

[RainStryke]

Even if a game can use more then 3GB of memory, it doesn`t mean that it cannot use less then 3GB of memory. As evilsofa point out, a game can preload textures into the available memory that the card have.

While you play the game, the textures can be swapped with other textures. I think the problem here is that 500MB of the 970 has low bandwidth, so any data stored on that area or data swapped on that area might slow things down causing framedrops and stuttering when used.

On those games, you`d probably be better off if Nvidia wouldn`t expose those 500MB to the game and instead just have 3.5GB available for the game instead.

Yeah, that would probably be the fix for the issue... for them to basically remove access to the partitioned memory. But the problem then will be lawsuits for marketing as a 4GB card and only delivering a 3.5GB card.

 

[Tamlin_WSGF]

Yeah, that would probably be the fix for the issue... for them to basically remove access to the partitioned memory. But the problem then will be lawsuits for marketing as a 4GB card and only delivering a 3.5GB card.

They might avoid that if the memory would be available for other tasks then gaming or just include an option to turn the partitioned memory off in the control panel (like they did with the compute option for Titan cards).

The 970`s are good cards and I would rather have 500MB less memory then to have stutter and FPS drops.

 

[RainStryke]

They might avoid that if the memory would be available for other tasks then gaming or just include an option to turn the partitioned memory off in the control panel (like they did with the compute option for Titan cards).

The 970`s are good cards and I would rather have 500MB less memory then to have stutter and FPS drops.

Agreed.

 

[n=1]

Even if a game can use more then 3GB of memory, it doesn`t mean that it cannot use less then 3GB of memory. As evilsofa point out, a game can preload textures into the available memory that the card have.

While you play the game, the textures can be swapped with other textures. I think the problem here is that 500MB of the 970 has low bandwidth, so any data stored on that area or data swapped on that area might slow things down causing framedrops and stuttering when used.

On those games, you`d probably be better off if Nvidia wouldn`t expose those 500MB to the game and instead just have 3.5GB available for the game instead.

Read somewhere that since nVidia knew about this issue all along, they made secret tweaks in the driver that avoided using the last 500MB of vram in the 970 as much as possible, and only tapped into the 500MB "reserve" if the game actually needed it.

On that point, as I've mentioned in this thread, I've seen Watch Dogs chew up 3.8GB vram at 1.78x DSR (1440p downsampled to 1080p), yet I did not experience any additional stuttering, so I'm inclined to agree with misterbobby that allocated =/= actively being used.

 

[EdZ]

There's a clear trend for everyone to chalk up every single performance issue they;re experience to "zomg the vRAMs!". This is not particularly helpful in determining if there is a problem, particularly if it is not accompanied with hard numbers. If you are experiencing stuttering, and do not have access to FCAT, then FRAPS can still grab a reasonable facsimile of frame render time in the majority of cases.

Remember NOT to use Shadowplay during testing, as that is allocating itself a chunk of vRAM to cache the framebuffer and encode (with some more vRAM needed for the encoding process) prior to passing the encoded video back over PCI-E.

 

[Tamlin_WSGF]

Read somewhere that since nVidia knew about this issue all along, they made secret tweaks in the driver that avoided using the last 500MB of vram in the 970 as much as possible, and only tapped into the 500MB "reserve" if the game actually needed it.

On that point, as I've mentioned in this thread, I've seen Watch Dogs chew up 3.8GB vram at 1.78x DSR (1440p downsampled to 1080p), yet I did not experience any additional stuttering, so I'm inclined to agree with misterbobby that allocated =/= actively being used.

You focus to much upon the word "allocated" I think.

If a certain amount of memory is allocated, it simply means its set aside for a certain task. The amount of memory can be used or it can be reserved. It can also be used for resources that are not accessed often or simply not at all, so that any stutter might not be so evident.

Problem is that there are cases where the game uses those 500MB of memory and users are experiencing stutter. According to Nvidia via PCPER, the first 3.5GB of memory is prioritized, so the issue shouldn`t be a problem in most cases, but since there are cases where this is an issue, it should be resolved.

I hope that Nvidia doesn´t use simply FPS delta to look into this problem as shown in the PCPER article, since its the stutter that people have problem with, not average FPS numbers. If 1-3% less Frames Per Second as shown in PCPERs article translate into 1-3% more Stutter Per Second its a problem.

 

[c3k]

Even if a game can use more then 3GB of memory, it doesn`t mean that it cannot use less then 3GB of memory. As evilsofa point out, a game can preload textures into the available memory that the card have.

While you play the game, the textures can be swapped with other textures. I think the problem here is that 500MB of the 970 has low bandwidth, so any data stored on that area or data swapped on that area might slow things down causing framedrops and stuttering when used.

On those games, you`d probably be better off if Nvidia wouldn`t expose those 500MB to the game and instead just have 3.5GB available for the game instead.

My bold.

I agree. I own a 970 and have seen some lockups. Is it caused by this? I don't know. If 500Mb of vram is inconsequential (I have dual 1920x1200 monitors), then allow me to disable it. Smoothness is paramount and second only to reliability. Eye candy comes a distant 3rd.

Ken

PS: and then offer me another 970, free, to make up for my pain.

 

[EdZ]

In those games, you`d probably be better off if Nvidia wouldn`t expose those 500MB to the game and instead just have 3.5GB available for the game instead.

That is exactly what is occurring at the moment. Then only time that last 500MB segment will be used in active operation is if a specific command will need to use more than 3.5GB to execute (e.g. you have two 1.2GB megatextures you need to blend into a third 1.2GB megatexture for a total operation usage of 3.6GB). Needless to say, this is a very uncommon occurrence. Otherwise, the 3.5GB is the portion of VRAM that is being used, and that 500MB is used to hold data that either isn't being used in an operation (e.g. cached textures that can be moved into the 3.5GB portion for active use) or data that doesn't require very high access rates (e.g. physics calculation results).

We also don't actually know the real drop access rate to this 500MB portion. The CUDA benchmark everyone is running does not actually access that portion of vRAM, and instead is paging to system memory (hence the bandwidth to that portion looking suspiciously close the the PCI-E bus bandwidth), because it was never designed to test this scenario.

 

[Quartz-1]

PCPer reports that Nvidia have responded.

Basically, the 970 suffers a major performance hit when using more than 3.5 GB RAM.

 

[GoldenTiger]

PCPer reports that Nvidia have responded.

Basically, the 970 suffers a major performance hit when using more than 3.5 GB RAM.

Actually, that is showing it only takes a tiny amount bigger a hit than the 980 does in performance when raising settings and using the higher VRAM amounts: the opposite of "major". What I want to see is FCAT data however as far as SLI and how it is affecting frametimes, because subjectively it is hitting frametimes badly from the small amount of playing I did trying that at 60+ fps but showing more than 3.5GB VRAM allocation. However, I'd much rather see objective measurements.

 

[Valtiel]

Actually, that is showing it only takes a tiny amount bigger a hit than the 980 does in performance when raising settings and using the higher VRAM amounts: the opposite of "major". What I want to see is FCAT data however as far as SLI and how it is affecting frametimes, because subjectively it is hitting frametimes badly from the small amount of playing I did trying that at 60+ fps but showing more than 3.5GB VRAM allocation. However, I'd much rather see objective measurements.

I agree. As far as I'm concerned, if the data from reliable independent testing shows that the last 500mb causes performance issues then Nvidia really needs to be held accountable for this.

However, right now it hasn't been 100% proven.

 

[MrMike]

Is comparing single cards really that useful?

I kind of have a hard time believing that a single 970 or 980 can be capped at near 4GB of RAM, have it not be texture caching (ex: Far Cry 4), and produce a smooth frame rate. I'm not saying that texture caching isn't useful, I'm just pointing out that the textures are just sitting there waiting to be used; they are not being actively displayed.

I think all we can do is wait for SLI numbers (and more data in general), as I don't think there's much debate that SLI Maxwell can push 4GB RAM effectively. The closest thing I can find is Guru3D's 980 SLI and 970 SLI reviews, but they're not doing anything to really push the VRAM.

At best, the current data suggests that the answer to the question of 3x970 vs 2x980 is 2x980.

 

[BlueStorm]

PCPer reports that Nvidia have responded.

Basically, the 970 suffers a major performance hit when using more than 3.5 GB RAM.

Absolutely wrong dude, sorry. They're showing that 980, which doesn't have this "special" type of memory config, drops performance by the same amount as the 970 when settings are turned up to use above 3.5GB. And the only reason performance dropped at all on both configs is that they changed resolution and settings... The GPU workload increased along with the memory usage amount.

Therefore it shows that the special memory configuration does not hurt game performance on the 970.

I suggest you be more careful about your assertions next time.

 

[BlueStorm]

I'm guessing you missed the part where Nvidia responded to the problems, mentioning that they partitioned .5GB of the memory and memory tests show that the performance of the partitioned memory is slower than molasses. They've cut corners and they are being called out on it... They have not directly responded to Nai's benchmark yet. Which is currently the proof behind the performance of the allocated memory.

Sorry this is wrong. Wow, so much misinformation in this thread. The Nai CUDA memory tests were *TESTING SOMETHING ELSE ENTIRELY*.

They tested allocations that were spilling to system memory, *NOT* the special section of video memory. Even Nai, the author of the test acknowledged this.

 

[BlueStorm]

Even if a game can use more then 3GB of memory, it doesn`t mean that it cannot use less then 3GB of memory. As evilsofa point out, a game can preload textures into the available memory that the card have.

While you play the game, the textures can be swapped with other textures. I think the problem here is that 500MB of the 970 has low bandwidth, so any data stored on that area or data swapped on that area might slow things down causing framedrops and stuttering when used.

On those games, you`d probably be better off if Nvidia wouldn`t expose those 500MB to the game and instead just have 3.5GB available for the game instead.

This is incorrect. Data stored in that area would be slower to access on every access, for every frame. Therefore it can't cause single frame stutter just to have data stored there. You would not see spikes or stutters, just that each frametime is slowed down equally by 1-2% overall.

The memory is much faster than system memory, so disabling the 512MB would be thoroughly retarded. Disabling it would *guarantee* that you get the performance numbers you saw in Nai's CUDA benchmark, since that's exactly what they measured... PCIE-bottlenecked system memory accesses.

The memory segment is much faster than that, and because important resources are prioritized not to be placed there, it's used as extra texture space for low priority textures, and would not affect performance much at all, as NVIDIAs data showed.

 

[RainStryke]

Sorry this is wrong. Wow, so much misinformation in this thread. The Nai CUDA memory tests were *TESTING SOMETHING ELSE ENTIRELY*.

They tested allocations that were spilling to system memory, *NOT* the special section of video memory. Even Nai, the author of the test acknowledged this.

Do you have a source to link on this? Or is this just here say?

 

[trick0502]

PCPer reports that Nvidia have responded.

Basically, the 970 suffers a major performance hit when using more than 3.5 GB RAM.

These are the like the most oddest settings and most random resolutions. It looks like they took their time to find benchmarks that fit there 1-3% claims.

When AMD had frame times all over the place when using crossfire NVIDIA developed fact. Where is there fcat testing?

We need some independent testing on this.

 

[BlueStorm]

Do you have a source to link on this? Or is this just here say?

http://www.overclock.net/t/1537725/...-gtx-970-3-5gb-memory-issue/180#post_23453567

Which links to:

https://translate.google.com/transl...5408&page=7&p=16912375#post16912375&sandbox=1

Which contains a post by Nai, the author of the widely circulated CUDA benchmark, saying it seems the last 512MB is not being used in his test, and it is instead going to system memory and therefore measuring PCIE bandwidth:

...
The benchmark measures "actually" not the DRAM bandwidth but the bandwidth of the "global memory" s in CUDA. The global memory in CUDA is a virtual memory space, which includes not only the DRAM, the GPU but also DRAM areas of the CPU. A virtual memory space is always distinguished by the fact that there is virtual memory addresses. The virtual memory addresses are translated to a memory access * somehow * in the actual physical memory addresses of the DRAM. The advantage of such virtual memory space, among other things, that it allows swapping of memory regions. Just such a retrieval at the DRAM, the CPU seems CUDA here to make the slump in the benchmark.

The benchmark allocates as many blocks of the global DRAM memories within the GPU as possible. Then it measures the read bandwidth within each of these blocks. The measurement is only as long as the allocated blocks really are in the DRAM GPU good. In this case, memory bandwidth is measured relatively accurately.

The problem however, is that the benchmark is not fully owns the DRAM, the GPU itself. After running in the background Windows and various programs, all also claim some of the DRAM of the GPU. However, the virtual memory space guarantees the benchmark that it () except for the area of ​​the primary surface (?) May be used for almost the entire 4 GiByte. Therefore, the allocation of non-DRAM proposes initially not be out until the upper limit of about 4 GiByte. However, the GPU must now begin to swap memory areas. That is, when the CUDA application is running, should in the DRAM of the GPU, the corresponding data of the CUDA application. Does the character process of Windows or another application should in DRAM GPU their data is.

The GPU seems swapping, according to my first simple investigations of this benchmark is to make simple associative. That is, at a physikalsichen address in the GPU DRAM always the same CUDA data or DirectX data must reside. This leads to conflicts that manifest themselves in the collapse of the bandwidth of the global memories in such conflict areas.

The benchmark attempts to reduce this effect by repeatedly requesting the data in each storage area successively. That is the first demand "should" cause the benchmark a page fault. The page fault "should" copy the data on the GPU Page from the DRAM from the CPU to the GPU DRAM. Then the other global memory accesses would be carried out with the DRAM bandwidth. So at least the assumption on my part.

Interesting way, the GPU does not behave as expected of me. This makes the GPU, the corresponding data does not seem to upload into the DRAM of the GPU, but again to request directly from the DRAM CPU with each memory access in a page fault in CUDA. Characterized the benchmark measures total in such cases, more or less the swapping behavior of CUDA, and not the DRAM bandwidth.
...

 

[BlueStorm]

These are the like the most oddest settings and most random resolutions. It looks like they took their time to find benchmarks that fit there 1-3% claims.

When AMD had frame times all over the place when using crossfire NVIDIA developed fact. Where is there fcat testing?

We need some independent testing on this.

No, they toggled settings to get memory in the sweet spots of just under 3.5GB and above it, but below 4GB enough to avoid normal video memory swapping that affect *every* GPU out there when you approach the dedicated video memory limit.

 

[polonyc2]

does this only effect people at extreme resolutions?...I'm at 1920 x 1200...this story gets stranger by the day...something obviously is not right with the 970 once you hit the 3.5GB VRAM point, but how much real world effect it has is what really matters to me

 

[RainStryke]

http://www.overclock.net/t/1537725/...-gtx-970-3-5gb-memory-issue/180#post_23453567

Which links to:

https://translate.google.com/transl...5408&page=7&p=16912375#post16912375&sandbox=1

Which contains a post by Nai, the author of the widely circulated CUDA benchmark, saying it seems the last 512MB is not being used in his test, and it is instead going to system memory and therefore measuring PCIE bandwidth:

Thanks for sharing that. The translation makes it very disorienting to read, it's hard to make assumptions in what he is trying to say since the sentences are so gargled with translation errors.

Maybe the micro stuttering issue can be fixed via driver refresh... Either way, it's too late for me, as i've returned them. But I hope people that are stuck with a GTX 970 see some kind of fix for the ones that are experiencing the stuttering issue.

 

[jwcalla]

There's a clear trend for everyone to chalk up every single performance issue they;re experience to "zomg the vRAMs!". This is not particularly helpful in determining if there is a problem, particularly if it is not accompanied with hard numbers. If you are experiencing stuttering, and do not have access to FCAT, then FRAPS can still grab a reasonable facsimile of frame render time in the majority of cases.

Exactly. Stuttering is an old problem that has a variety of possible sources but now everyone seems to think this is the cause of all stuttering issues for 970 owners.

What about those of us who don't have 970s? My GTX 570 has stuttering problems too (in some games). Can I blame it on segmented memory? Or could it be caused by something else? If in my case stuttering is caused by some other set of unrelated issues, is it possible that these same issues apply to a 970?


The way teh innernetz have flipped out over this... it's startling.

 

[Zinn]

It seems only games that are played on settings that actually "need" over 3.5 are having the issue.

When you purchase a video card that is advertised as having 4gb of memory on a 256 bit bus, it's reasonable to expect games that "need" over 3.5gb to run smoothly.

It's kind of painful watching you try to apologize for Nvidia in this thread. What they have done is blatant false advertising, and it makes no sense for you to defend them so heavily.

 

[misterbobby]

When you purchase a video card that is advertised as having 4gb of memory on a 256 bit bus, it's reasonable to expect games that "need" over 3.5gb to run smoothly.

It's kind of painful watching you try to apologize for Nvidia in this thread. What they have done is blatant false advertising, and it makes no sense for you to defend them so heavily.

You need to pay more attention to the context of my posts. I have even flat out said that I am not defending them. All I am saying is that a game that actually needs more than 3.5 GB will not be playable anyway regardless of this issue. With SLI this will be a much bigger issue though especially down the road so I am glad all the tech sites are bringing this to light so Nvidia has to answer for this.

 

[BlueStorm]

Thanks for sharing that. The translation makes it very disorienting to read, it's hard to make assumptions in what he is trying to say since the sentences are so gargled with translation errors.

Maybe the micro stuttering issue can be fixed via driver refresh... Either way, it's too late for me, as i've returned them. But I hope people that are stuck with a GTX 970 see some kind of fix for the ones that are experiencing the stuttering issue.

I assure you these key couple of sentences are very clear:

Interesting way, the GPU does not behave as expected of me. This makes the GPU, the corresponding data does not seem to upload into the DRAM of the GPU, but again to request directly from the DRAM CPU with each memory access in a page fault in CUDA. Characterized the benchmark measures total in such cases, more or less the swapping behavior of CUDA, and not the DRAM bandwidth.

He is saying interestingly, the GPU doesn't behave as he expected. The data he uploads to the video memory (DRAM) on the GPU to measure bandwidth seems to always be coming from (requested from) CPU system memory (CPU DRAM). Therefore, in these cases the benchmark measures this swapping behavior (CUDA accessing slow CPU system memory) and not the GPU DRAM bandwidth.

This means the amounts coming from system memory above 3.5GB are likely a bug in how CUDA is choosing to allocate memory. CUDA (not graphics or games) is mistakenly ignoring the top 512MB. Whatever it is, it is specific to CUDA.

Maybe it is easier for me to interpret because I am experienced with the system architecture and how the various types of memory work together.

The measured bandwidth amounts people post in this situation are always in line with PCIE bandwidth, which is theoretically 32GB/s, but in reality PCIE always achieves less than that (bus/signaling efficiency).

System memory bandwidth for new CPUs are well above that amount of bandwidth now, but you'll notice that despite this fact, the numbers they report are always lower.... Because PCIE is a bottleneck that sits between the GPU and the CPU/system memory.

Also he mentioned that his test is just allocating memory on the GPU and keeping track of how much HIS APPLICATION is using... It is not measuring how much is actually consumed on the GPU. So I don't believe the "amount consumed" incorporates things like Windows Aero, which is why everyone is saying you need to run Nai's benchmark headless. It also doesn't incorporate any baseline memory that is used by the driver simply to make the GPU work, though this amount should be fairly small.

 

[RainStryke]

I assure you these key couple of sentences are very clear:



He is saying interestingly, the GPU doesn't behave as he expected. The data he uploads to the video memory (DRAM) on the GPU to measure bandwidth seems to always be coming from (requested from) CPU system memory (CPU DRAM). Therefore, in these cases the benchmark measures this swapping behavior (CUDA accessing slow CPU system memory) and not the GPU DRAM bandwidth.

This means the amounts coming from system memory above 3.5GB are likely a bug in how CUDA is choosing to allocate memory. CUDA (not graphics or games) is mistakenly ignoring the top 512MB. Whatever it is, it is specific to CUDA.

Maybe it is easier for me to interpret because I am experienced with the system architecture and how the various types of memory work together.

The measured bandwidth amounts people post in this situation are always in line with PCIE bandwidth, which is theoretically 32GB/s, but in reality PCIE always achieves less than that (bus/signaling efficiency).

System memory bandwidth for new CPUs are well above that amount of bandwidth now, but you'll notice that despite this fact, the numbers they report are always lower.... Because PCIE is a bottleneck that sits between the GPU and the CPU/system memory.

Also he mentioned that his test is just allocating memory on the GPU and keeping track of how much HIS APPLICATION is using... It is not measuring how much is actually consumed on the GPU. So I don't believe the "amount consumed" incorporates things like Windows Aero, which is why everyone is saying you need to run Nai's benchmark headless. It also doesn't incorporate any baseline memory that is used by the driver simply to make the GPU work, though this amount should be fairly small.

It's all about perception on how you look at it... I went in looking at it with your point of view and agreed at first... but I went over and read it twice over... I'm not so sure that's exactly what he is saying. A lot of those translations can make out something entirely different that what he is trying to say. I'm not saying you are wrong... I'm just saying, I just don't see it cut and dry.

Exactly. Stuttering is an old problem that has a variety of possible sources but now everyone seems to think this is the cause of all stuttering issues for 970 owners.

What about those of us who don't have 970s? My GTX 570 has stuttering problems too (in some games). Can I blame it on segmented memory? Or could it be caused by something else? If in my case stuttering is caused by some other set of unrelated issues, is it possible that these same issues apply to a 970?


The way teh innernetz have flipped out over this... it's startling.

Referring to the GTX 970 customers as "teh innernetz" is annoying and relating it to your GTX 570 issue is not the same. This has blown up because Nvidia rushed these cards out and tried to patch up the issue with a quick fix.

They released a GTX 970 with dual 6 pin power. TONS of customers were complaining about the video cards down clocking under stress because the GPU isn't getting the power that is needed, especially the customers that spent extra on the factory overclocked versions. So they release a new version of the GTX 970 with a 8 pin and a 6 pin connector to allow more power phases to the GPU to alleviate the power throttle issue.

Now, we have people that can't fully utilize the 4GB of memory without it stuttering. It could be something to do with the new power phases... It could be something with quite a few batches of VRAM chips... or it could be something with the inductor coil whine that we are hearing so much about. That coil whine could be a sign of cheap inductors or some other poor quality counter part that is causing it. I really don't know... but I do know one thing... I hate having to troubleshoot my hardware and tweak things just to get it to work as intended when I spent a considerable amount of cash to play my video games and have fun... Not to be stressed out wondering if my hardware is going to get a firmware, driver or replacement fix.

Looks like the red team just dropped prices again across the board. You can now get a Sapphire Tri-X or MSI Lightening R9 290X for $350. They are definitely taking advantage of this situation. My Sapphire R9 290X Tri-X pretty much stays at the same temps as the GTX 970's I had... The drivers and performance of this card is solid. I'm talking with my wallet now. Nvidia may be in my future... later on, but for now, they are off my spending list.