GTX-970 VRAM technology limitations (or any video cards VRAM).

rower30

Distinguished
Dec 16, 2002
264
0
18,790
VRAM technology implications.
This “question” is NOT intended to drag the legal aspects of “marketing truths” into the discussion, so much as to review the GTX-970 hardware as it really is, and we know that now, and what it means.
-
I don’t see too much in the way of “truthful” technical answers to the current hardware configuration. No, I don’t know the exact performance truth other than my RECENTLY purchased GTX-970 works extremely well.
-
Can people who KNOW the right answers chime-in? DO NOT comment on the darn marketing goof, go elsewhere, please. I'm going to sit back and learn, so don't expect responses per say. I can't listen with my mouth open. I want to know the VRAM's configuration limits.

- My AMD 6870 ran with I Gig VRAM at @ 1080P on a 24” LCD with a single notch down or two on most games and was never, "a stuttering unplayable mess”. Why not, as it certainly hit the 1 Gig VRAM ceiling?
- When a card does hit the VRAM ceiling, it seems ANY card forced to go to the system RAM will see a performance hit, although I don’t see this on my AMD 6870 @ 1080P. The GTX-970, and I see it as a 3.5 Gig card with a “soft” hit to the last 0.512 Gig verses a hard hit to system memory above 3.5 Gig. How is this cache technology a performance issue, really? It seems the best solution with this GPU architecture cut-down from the GTX-980 to design in a soft hit and use this memory, than to make the card 3.5 Gig and delete it. You can CHANGE the quality setting to avoid the above 3.5 Gig hit if you like, or, as we do with current “full speed” cards, adjust to avoid the hit to system memory at all.
- In SLI, you only see the memory from the master controller card, so you get 4 Gig even with two GTX-980. And, those using SLI seem to want to run two or three monitors with ½ to 1/3 the single monitor VRAM. These card’s aren’t magic, and you will see VRAM limitations not apparent on single monitor solutions, requiring reduced quality setting. The 980 will hit the VRAM limit and then turn into a “stuttering unplayable mess”, right? OK, the 970 does so slightly under the 980, but where’s the beef in this argument for the price of the card? The 970 is $200.00 cheaper, so WHY would it NOT have limitations not seen in the 980?
- Future proof? Me, I see this as crazy. I buy for TODAY as four or five years from now, even if I had 16 Gig of VRAM, the GPU will be so much different that my “future proof” card will be obsolete. I KNOW that as games go forward I’ll also adjust features to suit. I see no real reason to STOP GPU development so your 16 Gig of 2015 VRAM is good in the year 2020, as though that's the only progress to be made the next 5 years. Do you seriously believe this future proof issue? High-end hobbyists will ditch a 980 in a heartbeat for one significant GPU change VRAM be damned.
- These video cards have THOUSANDS of specifications not on the box. Is THAT a conspiracy to performance? No, the end game (pun intended) is the overall performance to game play. THAT is what I looked at, and NOT a spec in and of itself. I use the card, not a spec. How is the current tested cards performance not accurate to tested data?
-
So, the BIG question is, is a 3.5 Gig card with a soft hit to .512 Gig local cache and THEN to system RAM worse than a 3.5 Gig card that hits system RAM straight away?
-
I can’t see the logic that the latter is better than the former. And, about every card out there except the GTX-970 hit system memory right after the VRAM limit. To avoid the memory latency, you need to reduce some settings, right? I see no complaints about that. My 1 Gig 6870 surely would have seen the issue…and I don’t see, “an unplayable stuttering mess”. It was amazingly good, actually, for a four+ year old card.
-
So in the end, it seems you have a great 3.5 Gig card that needs maybe a tweak or so of visuals to see good game play just before you might need them on the 980. How is that not expected for $200.00 less? How does a soft local .512 cache make the GTX-870 a bad card? What is the “problem” with the “feature” from a pure technical standpoint?

OK, chew me up.
 


So here's the deal.

Most modern GPUs are marketed as having a large, seemingly uniform memory bus. For example, the Radeon R9-290X is marketed as having a 512 bit bus, and the GeForce 970/980 are marketed has having a 256 bit memory bus. This is no more true than were the Intel i7-4790K marketed as having a 128 bit memory bus or the i7-5960x having a 256 bit bus.

The marketed number is obtained by concatenating the width of multiple independent buses that are attached to one or more memory controllers. Dual-channel, triple-channel, and quad-channel are terms well known to CPU enthusiasts, so marketing departments don't need to outright make things up. However, GPU microarchitecture is not as well understood by the masses (or at least to the extent that some people think that they understand them) as CPU microarchitecture. As a result, marketing departments might have a hard time selling a GPU with four memory channels or eight memory channels.

All GDDR5 SGRAM chips have a native a 32-bit IO interface, so 32-bit single-rank and 64-bit single-rank channels are commonly used for maximum data rates. NVidia's Maxwell microarchitecture has four memory controllers, each controlling a pair of 32-bit buses that each have a 4Gib GDDR5 SGRAM chip (or a pair of 2Gib chips) attached to it. Each memory controller is connected to a Graphics Processing Cluster (GPC) and each GPC has four Streaming Multiprocessors (SMMs) contained within it. A crossbar switch joins the GPCs and memory controllers together, allowing any SMM to access any L2 cache memory attached to the GPU and through the cache access the memory controllers.

On the GTX970, three out of the sixteen SMMs are disabled from any of the GPCs. This is done for marketing and yield reasons. In addition to this, eight ROPs and one chunk of L2 cache (256KiB) are disabled as well.
The disabled ROP/L2 cache component creates a bottleneck. Each memory controller (two 32-bit channels) has a pair of L2 cache blocks sitting between it and the crossbar. There is another switch between the cache and the memory controller, but by disabling one of the two caches, that memory controller's bandwidth to the crossbar switch was crippled. In order to maximize throughput, the channel behind the disabled ROP/L2 cache was weighted lower in memory allocation priority. It's still there, it can still be used, and it's still faster than spilling over to shared system memory, but it can't transfers as much data as the other memory controllers due to the crippled front end.
 

rower30

Distinguished
Dec 16, 2002
264
0
18,790
And when ANY video card hits the VRAM limit what happens? Even the "unbroken" ones?

It looks like the GTX-970 is simply a 3.5 Gig VRAM video card...adjust to suit. No different than when I set my 1 Gig card to limit VRAM overload, yes? or when you adjust your GTX-980 just the same after you load all those texture mods.

I simply do not see the issue in the GTX-970 card's actual design decision, and again, the marketing part be damned for now. Would the "truth" sway may decision on what card to buy? Not at all, and it didn't. All card's that aren't ten tenths cards are technically crippled (to use a loaded term) but MORE than suitable for different levels of performance. GTX-970 and 1080P? Sign me up.

The marketing stuff is irrelavent to me technically. It should be to you too as the TESTS are what matter. Those test were there to verify the performance when cards were purchased. The GTX-970 does not show distress (stuttering mess) as a 3.5 Gig card when set-up properly and for the VRAM it has. Yes, there is a small deficite to the 980 in texture performance I grant you that. But "as is" the card works very well.

An R9 280X or even an R9 290 (the OC R9 290X's are tough to beat) have trouble keeping up with FPS to a moderately overclocked GTX-970 so we trade theoretical VRAM texture for FPS. I say theoretical because I have no clue how NVIDIA verses AMD manage memory. Are the two vendors 4 Gig VRAM cards really the same allocation for textures? I don't know what the answer is other than a decisoin as to what best suits you. I'd rather move right along with less texture detail later in a card's life. A slower card with more VRAM? I don't think so.
 
And when ANY video card hits the VRAM limit what happens? Even the "unbroken" ones?

That depends. With Windows Vista, Microsoft introduced the Windows Display Driver Model version 1.0. One of the features that WDDM introduced is virtualized video memory. The graphics driver can swap adapter memory out to physical system memory in the same way that the operating system's memory manager can page physical memory out to the hard disk. NVidia has simply introduced a middle man in the form of a lower priority pool. If the adapter needs more than 3.5GiB it starts to allocate from the 512MiB low priority pool which shares bandwidth with 512MiB of the regular priority pool. If the full 4GiB is exhausted, it starts paging to system memory.

It looks like the GTX-970 is simply a 3.5 Gig VRAM video card...adjust to suit. No different than when I set my 1 Gig card to limit VRAM overload, yes? or when you adjust your GTX-980 just the same after you load all those texture mods.

It's a full 4GiB card, and I guarantee that if NVidia tried hard enough they could come up with a memory management scheme that made better use of the crippled memory controller. However, I doubt that it's worth doing so for a single card that already performs very well.

I simply do not see the issue in the GTX-970 card's actual design decision, and again, the marketing part be damned for now. Would the "truth" sway may decision on what card to buy? Not at all, and it didn't. All card's that aren't ten tenths cards are technically crippled (to use a loaded term) but MORE than suitable for different levels of performance. GTX-970 and 1080P? Sign me up.

It's understandable that you don't see the logic in the design decision, most people wouldn't. I'm an engineer so I deal with these kinds of decisions all the time. The simple explanation is that although all GM204 chips are designed equally, they are not all manufactured equally. Defects are present in all integrated circuits and sometimes these defects are severe enough to render a portion of the product unusable. Defects are measured in defects per square centimeter, so it should be apparent that as the size of an integrated circuit increases, the number of defects contained within it increases as well. Furthermore, it should also be apparent that the defects are more likely to occur in larger components of the chip, especially caches. This is why all major designers (Intel, AMD, NVidia, IBM, Samsung, Qualcomm, etc...) have adopted designs that allow them to selectively disable portions of a chip. By disabling a portion of a chip, they can sell it as a lower tiered product rather than discard it completely.

The i7-3930k and i7-3960x are actually 8 core microprocessors with 20MiB of L3 cache. Each of these processors has two cores disabled and 8/5MiB of L3 cache disabled respectively. Which sections are disabled is assessed on a per-chip basis. Flawless variants of these chips are sold under the Xeon branding as the E5-2650+ series.

The marketing stuff is irrelavent to me technically. It should be to you too as the TESTS are what matter. Those test were there to verify the performance when cards were purchased. The GTX-970 does not show distress (stuttering mess) as a 3.5 Gig card when set-up properly and for the VRAM it has. Yes, there is a small deficite to the 980 in texture performance I grant you that. But "as is" the card works very well.

Right. NVidia didn't come out and cripple the cards after they were released. The day one benchmarks are still valid; nothing has changed there.

An R9 280X or even an R9 290 (the OC R9 290X's are tough to beat) have trouble keeping up with FPS to a moderately overclocked GTX-970 so we trade theoretical VRAM texture for FPS. I say theoretical because I have no clue how NVIDIA verses AMD manage memory. Are the two vendors 4 Gig VRAM cards really the same allocation for textures? I don't know what the answer is other than a decisoin as to what best suits you. I'd rather move right along with less texture detail later in a card's life. A slower card with more VRAM? I don't think so

The SGRAM is used for far more than just textures. It stores all kinds of buffers including meshes, textures in multiple transformed formats, frames, depth and stencil buffers, etc.. It also stores compiled shader programs, geometry information and loads of other fun stuff. At a high level, the GPUs from all major vendors are remarkably similar.
 

rower30

Distinguished
Dec 16, 2002
264
0
18,790
OK, excellent.

I'm an engineer, too but when I act like one I post your FIRST response...which doesn't really answer the issues for a laymen to understand. It answers technically how it got there but not where that actually is. Your SECOND post is excellent, and does put the data down so it is understandable to the masses. The closer you can get to a "yes" and "no" answer the better. "Yes" the car does 60 MPH. The first response discussed the BMEP of the engine, CD of the chassis and gear ratio matching to the engine's torque peak and on and on. OK, but it didn't answer the question, how fast is it?

Is the GTX-970 WORSE than 3.5 Gig VRAM full speed card? NO, it isn't.
Is the GTX-970 WORSE than 4.0 Gig VRAM full speed card? YES, it is.
Is the GTX-970 a good card? Depends (oops!!).
Cost - You can look UP at the GTX-980 or DOWN to the R9-290X.
Efficiency - There is nowhere to look but at the GTX-970 anywhere near the price.
Features set - NVIDIA and AMD both have different feature sets.
Multi panel usage - Here I feel most SHOULD go for FULL speed VRAM and large amounts and not even
mess with 4 Gig cards if your texture VRAM usage stories are correct. Are they?

A R9 280X 3.0 Gig VRAM card and the GTX-970 3.5 Gig VRAM card will be hard to distinguish in single panel usage, except that the "extra" 0.5 Gig full speed VRAM and the "soft" hit 0.5 Gig VRAM above that are a slight advantage on the GTX-970 before hitting system memory. Power use is no contest with the newer NVIDIA GPU.

Marketing goof aside, from and Engineer's perspective...and for informed buyers of the GTX-970, it is a well done card. View it as a 3.5 Gig card and compare from there.

If your power supply needs are addressed, an R9 290X is a good price comparison with 8 Gig memory. Two of those for SLI will need a HONKING power supply! But, it is what you multi panel users should be looking at for bang for the buck. And, the card do seem quiet. My AMD 6870 was a peach on noise. So NVIDIA handed you a sweet deal on the R9 290X price! If I had 4K, I'd be hard pressed to look at two GTX-980 with just 4 Gig VRAM. Buy two 8 Gig VRAM R9 290X for $625.00 and get a HUGE power supply for the price of just two GTX-980's with 4 Gig VRAM and STILL have money left over? That's a tough deal to pass up.

For single panel 1080P users, and with more modest system builds, the GTX-970 is a fantastic and efficient card. I'm going to skip above 1080P until designs are out that more suitably match GPU and VRAM to them as a system. As is, it's a wild guess with somewhat poor system optimixation at best. Yes, 4K will run. It will "just" run the game and run plenty of $$$. I'm happy with 1080P...FINALLY done VERY well with a single GTX-970 and 60 Hz monitor Vsynced to 75 Hz ( I overclocked the S24D590L). But THAT took awhile!

I will upgrade to an adaptive sync (one or both in one panel, who knows?), high resolution panel and with better a MATCHED video card than anything out now. I work in four year increments and even longer on the LCD panel as they took SO LONG to get decent for a reasonable price with an optimized sinmgle card solution. Can't be happier with the GTX-970 and SAMSUNG S24D590L panel. It's what was needed all along. Its here NOW optimized, I have a small power supply and one panel. I grabbed the GTX-970 an don't regret it one "bit".
 

rower30

Distinguished
Dec 16, 2002
264
0
18,790
iamlegend
February 8, 2015 3:22:18 PM
-
My point is, many feel there is something "wrong" with the GTX-970 with 3.5 Gig VRAM. Well, ANY card will run out of VRAM and simple adjustment to the graphics settings are easy, and inevitable depending on your monitor size, games used and plenty of other stuff.
-
WHY would a GTX-970 be any worse in game play than a R9 290X with 3.5 Gig VRAM, for instance? No one anywhere is saying that the R9 280 / 290X are an "unplayable stuttering mess". I used a 1 Gig AMD 6870 and it plays FarCry 3 pefectly fine (no stuttering or screen tearing) with a few graphical tweaks. The downside? Same as always, the newest cards have better detail and picture quality and CAN run tougher games well. Energy efficiency is nice, but hardly a deal breaker in my book, at least.
-
I find it odd that some owners of the GTX-970 all of a sudden claim the card is an utter disaster in game play out of the blue. Sure, in SLI it demands MORE then even 4 Gig of VRAM, so WHY get a 4 Gig card at all? I pointed out that the 2 x GTX-980 isn't a deal compared to the 2 x R9-290X in cross-fire with 8 Gig of VRAM, and this sells for LESS and even with a new 1000 watt power supply! 3.5 Gig to 4 Gig is hardly a "solution" nor a disaster in single panel systems when used properly. And, 8 Gig is more a solution to multi-panel support where to me 4 Gig is NOT.
-
The cards aren't magic. From a pure technical standpoint, I just can't see the reasoning the 4 Gig full speed VRAM is a "solution" over 3.5 Gig VRAM. 4 is "better" than 3.5...but is it a practical better? No, only technically. People need to understand video cards workings and not expect a 4 Gig GTX-980 to be this magical fix with a scant 500K more memory in 3 panel SLI. It won't. You need 8 Gig AT LEAST to cover that 4K triple panel in current games with added texture files.
-
The same hold true for 1080P. The GTX-970 has plenty of VRAM for most any game @ 1080P, and those that need more, the 4 Gig solution is still not really enough. I run the dreaded ACU (Assassin's Creed Unity) game maxed out on 1920 x1080P on the GTX-970 with an i7-870 CPU that isn't a modern power house and it runs flawlessly in smooth graphical detail. Well, until the game engine does what it does every fifteen minutes, anyway (locks up or pops to the desktop or refuses to start ETC) but that's not the card. It worked well till the 11/2014 "update".
-
Those that think the GTX-980 4 Gig will "fix" problems in SLI won't be happy. All those added textures will eventually overload the 4 Gig VRAM. WHY are "new" multi-panel users not getting the excellent 8 Gig AMD cards in Cross-fire that add all those textures? The GTX-970 / 980 brought those R9 cards down to your price range and then some.
-
In four years, even 8 Gig of VRAM WILL not be useful amongst all the other improvements to the GPU, so why pretend to future proof?


 

iamlegend

Admirable
Are you the savior of 970? Nobody argues with you. Facts are facts, 970 is a beast in 1080p hands down. Will you buy the 970 if you expect to game in 1080p? Yes, no, maybe?

Yes, I will play only on 1080p my whole life.
No, I will upgrade into 1440p, believing that this is a strong card.
Maybe, 1440p, 4k or SLI.

I will play only at 1080p, which GPU should I choose, the 280X or 970?

Get the 280X, it can max out all the games at 1080p
Get the 970, you can SLI it in the future to go with high res monitor.

These are facts, these are what we had believed. These situations happened.

Imagine if it were marketed to 3.5GB VRAM, do you think it will go head to head with the 290/290X? Yes, no, maybe?

Yes, it will for 1080p or 1440p at least.
No, it will not considering I will get 1440p Ultra settings.
Maybe, I don`t know.

Yes, I know it`s vague.
 

rower30

Distinguished
Dec 16, 2002
264
0
18,790


- Nothing vauge about it. Texture settings are, and always will be, made of choices. My AMD 6870 1 Gig maxed out in it's day and runs very well if set down to medium today. Nothing there surprised me at all. Yes, I knew out a few years what to expect, too. The new 1080P monitor meant a new card to set TODAY'S games at ultra settings. And yes, you will work down from there going forward. This GTX-970 will wear the future no different than my 6870 did.

The issue is, multi panel 4K is broken NOW, and with todays cards. This happens from time to time where stuff isn't lining up. But it isn't a mystery where the issue is.