GeForce GTX 970: Correcting The Specs & Exploring Memory Allocation

Status
Not open for further replies.
^^ I don't think so. Detailed low level technical specs like these are sent only to reviewers and its an easy enough slip up to make and not notice if the marketing teams and the core technical teams are not in sync. Just think, how many GPU design engineers would go to Anandtech and dozen other review sites, read the listed specs of the GPU that they have worked on and notice that one particular detail is wrong in all the sites. If anything, all that sort of reading would have been done by the marketing teams only.

In any case...

1. nVidia had nowhere advertised the detailed specs like ROP count of the GPU on their website for public viewing. I am pretty sure most people would not have relied on the ROP spec to decide whether to buy it. They would have relied on the actual benchmarks.

2. The card is supposed to have 4 GB VRAM and a theoretical peak memory bandwidth of 200+ Gbps and neither of them are incorrect. I don't see any basis to sue nVidia over this as none of their public advertising is wrong. It won't stop the people from trying, but not sure how fruitful any such attempt will be.

3. Even though this is the first time they made a GPU with ability to switch off ROPs, it is apparently not the first time they have used asymmetric memory configurations since some of the other GPU series have also used segmented memory partitions. Nobody reported such problems earlier

4. Maxing out the 4 GB VRAM requires ultra high quality textures and ultra high resolutions like 4k where a GPU like 980 leave alone 970 would loose a lot of performance. Memory availability should become less of an issue than GPU power by then. I don't think most single GPU users should even be bothered by it. The question is whether SLI users have any impact.

5. GTX970 costs less than 60% the price of the GTX980 and offers performance somewhere around 80% and even higher in many cases. So you already got more than fair value for the money.

Overall, I think it is an issue which is being blown out of proportion. Also, it is still not clear that this is the root of the performance losses that only a few people have been noticing. As people have posted across various forums, there are numerous GTX970 users who have not faced by performance issue and there are other older GPUs which used similar memory addressing and nobody has pinpointed a performance issue because of it.

I think it might more likely a bug in the drivers for 970 because of its segmented memory that is causing issues for certain users under certain conditions.
 
  • Like
Reactions: Crapmypants
Not really, the specs are available for all to see especially a major one like 56 ROPs vs 64. Don't fall for the damage control BS they are spewing.
Just because some people don't look into all the details doesn't mean you can falsely advertise, plus i doubt a company on the level of Nvidia doesn't know of problems on this level.
If you follow tech in general you'd know the 980 really isn't the top end card, it's an upper mid-ranger being sold for premium pricing, they had to shave some of the performance off the 970 so they could milk the fanboys more.
I agree the issue won't even be noticed by most people since i don't know of any title that will make utilise 4gigs at 1080p but there will eventually be a problem down the line.

This is exactly the tech equivalent of the Triumph street triple debacle in India where they exagerrated the power specs but Triumph came out with an apology and offered 1L refund to all buyers, what'll Nvidia do in this case ?
I don't think it's being blown out of proportion, they've done shady things in the past which is why i'm not giving them the benefit of the doubt.
 
  • Like
Reactions: RD274
^^ Can you show me any place where where nVidia has officially advertised ROP count either on their website or on a Graphics card box? I have a Gigabyte G1 GTX 980 and specs like ROP count are not mentioned anywhere on the box or manuals. I guess this would be the same for any GTX 970 boxes as well or for that matter any other GPU.

In order to call it false advertising, nVidia needs to have mentioned that spec somewhere in their official advertisement materials or in official product specs intended for the customer.

In fact, I don't see why nVidia would lie about the specs like ROP count which most customers don't use at all in their buying decisions. I don't think nVidia is lying about it at release because there is no reason for them to lie about it. I also don't think they are lying about it now as part of damage control because they could have easily come up with something much better if they really intended to lie for damage control.

Also, not sure why you would call a GTX 980 a mid-ranger considering it is the top GPU right now as far as single GPU goes. In fact 970 is the upper mid-ranger. GTX970's are technically the same die as GTX980. When they go about making the GTX980's, the yield they get is not 100% perfect. The imperfect parts many not match the 980 spec either because they have faults or because they are not stable at the clocks that 980s are supposed to run. These faulty parts have units disabled and clocked lower to make them stable and that's how you get the GTX 970. Without this sort of strategy, 970 would not exist at all and GTX980 would be much much costlier.

As I mentioned earlier, there are other older GPUs which used similar memory interfacing patterns and people have not reported issues. So I think there might be something else wrong for the users who are affected.
 
  • Like
Reactions: Crapmypants
Lord Nemesis said:
Can you show me any place where where nVidia has officially advertised ROP count either on their website or on a Graphics card box? I have a Gigabyte G1 GTX 980 and specs like ROP count are not mentioned anywhere on the box or manuals. I guess this would be the same for any GTX 970 boxes as well or for that matter any other GPU.
Click to expand...

GPU boxes usually lack any kind of in-depth info of the card and are mostly littered with marketing BS but the spec sheet is usually displayed in reviews and gives a rough idea of how it compares to the rest of the cards in the family and how badly it has been nerfed. That's how the card specs are usually advertised, you aren't gonna see ads for gpus in a newspaper. It's up to the consumer to be well informed before making a purchase, sure you can say no one looks at ROPs and this and that but these same people are the ones who also buy rebadged cards over generations without taking a look.

Nvidia just thought no one would notice but the enthusiasts proved them otherwise.
Sure the card has 4GB vram and bandwidth of 200GBps+ as you mentioned earlier but have you seen how it drops drastically when the memory usage crosses 3.5 gigs ?
Suddenly they are throwing out these block diagrams explaning how it works, any chip designer would have understood or tested for this behavior before the card came out on the market.

Lord Nemesis said:
Also, not sure why you would call a GTX 980 a mid-ranger considering it is the top GPU right now as far as single GPU goes. In fact 970 is the upper mid-ranger. GTX970's are technically the same die as GTX980.
Click to expand...

It's all in the code name, both the 970 and 980 are based on the GM204. The GXx04 is usually reserved for the mid-uppermid tier cards e.g., GK104 (760,760Ti,770), GF114 (560, 560Ti). The top of the line cards are mostly based on Gxxx0 based cards like GK110 (780, 780Ti, Titan).
Like i already mentioned if you follow hardware news you already know Nvidia is holding onto the real high-end cards, namely the GM200 (980 should have been here,980Ti,Titan) until AMD makes a move. See how they pull out their trump card once the R3XX is revealed around the April-June period.
It's the same reason why the GTX 960 is being called a fail card since it's based on the GM206 and being marketed as a mid-ranger since GXx06 was mainly reserved for lower mid-range cards like the 550Ti and 650Ti.
But hey, they made a good card in the GM204, slapped on the 980 sticker and sold it for $200 more for a 20% performance bump even though it's not the best they had.
 
  • Like
Reactions: Crapmypants, RD274 and Fenix
^^ Yes, only the material given to reviewers will have this sort of details and it does not really count as general purpose advertising material and I don't think any body can sue nVidia over incorrect specs sent to reviewers. Also tell me one reason why nVidia would like to intentionally send wrong about something like ROP count. The only thing that counts for customers in the end is performance which is measured through various benchmarks and tests and 970 performs well where. In fact, if they want to skew the specs intentionally, it would actually be beneficial for them to under quote the specs than bloat them as that would create an impression of the architecture being far better than it is. So I repeat again that nVidia has had no logical reason to deliberately give wrong information about ROP count to the reviewers.

In any case, you are missing the main point. Even though this is apparently the first architecture where they have the ability to partially enable/disable ROPs, non standard memory configuration itself is nothing new. It has been around and people have already proven that its there with older GPUs as well. If no consumer noticed it so far with older GPUs, I don't see why it would be a problem only for GTX 970.

If you ask me, I think people on forums tech jumped the gun on blaming the segmented memory architecture for the performance issues without knowing enough details to really pin point it as the culprit. If its really a architecture issue, everybody having a GTX970 should be affected. But that is not the case. Regardless of what Nai's benchmarking tool shows, Only a few people have reported issues when running actual games with VRAM usage above 3.5 GB. Yes, it affects performance a bit and there is no denying that, but then, if it is so brazenly visible, it should have been reported for every GTX970 out there and a dozen other GPUs that have used similar memory configurations in the past.

My opinion is that the performance problem is somewhere else and nVidia is yet to locate the root cause themselves or yet to reveal it to public.

Lastly, as far as the code names go, I don't think it matters what codes they have been using internally. In the market, GTX980 is the fastest Single GPU right now and nVidia is charging a premium accordingly. nVidia may come up with a GTX980 Ti, GTX985 or GTX990 in future, but till then GTX980 is their flagship high end GPU in the market.

Also FYI, the nomenclature you pointed is not standard the way you are thinking it is. The Kepler series (6xx series) has the codename GK104 for GTX680, 670 and 660. 7xx series were a combination of Keplers and Maxwells with the Maxwell being used for the newer lower end GPUs and denoted by the GMxxx codes. The GTX780 was a refresh of Kepler GK104 and used the code GK110.

So it really makes sense that the first Maxwell high end is denoted by GM204. The subsequent refresh may be named GM110. In this this may be in a newer series than GTX9xx.
 
Last edited:
  • Like
Reactions: Crapmypants and Fenix
Ultimately it boils down to the fact that people are feeling cheated. I PAIDE FOR 4GB VRAM AND 64ROP, HOW CAN THEY NOW SAY J/K TIS ONLY 3.5GB FAST+0.5GB SLOW VRAM AND LESSER ROP HURR DURR.

But nevermind all this confusion, what really bums me here is that this card has been on sale for a while and the reviews have been around for even longer, so what took Nvidia so long to clarify this? This wasn't some small misconception. Every review site stated these specs confidently so who would doubt them when they all stated the exact same numbers. Clearly someone in the company had to know what was up, so rather than trying to be good guys and clear this up before it could risk becoming a fault, they let it slide and hoped no one would notice. Its only now that the reviewers are being blamed for not accurately testing the hardware and that Nvidia is taking flak for not 'lying' about specs that these issues are finally being addressed. And ever since the Kingston and PNY issue, people are desperate to avoid being lied to because honestly, performance aside who wants to have a hardware product that everyone knows is flawed? And the real irony in all this, is that the benchmarks and scores remain the same regardless of the debacle ... its just bad PR. On one side, Nvidia is lucky it took so long to come in the open and didn't risk sales. But on the other hand people are going to be sceptical on their next batches of cards, especially the fickle minded enthusiasts.

Nvidia should just give all GTX970 owners affected a game and let it pass. I honestly dont see a better PR stratergy for them.
 
  • Like
Reactions: psyph3r
I think most people paid for a card that performs at 80% or more but costs less than 60% the cost of GTX980, and not for getting 64 ROPs same as 980. People are already getting more than they paid for.

If numbers comparisons of ROP count etc are so important for people in making a buying decision, Would they be fine paying the same cost for a card with 128 ROPs, but performs at half that of 970? People care about performance more than the specs. This is exactly why these low level specs don't appear anywhere on your graphics card boxes. specs like clock speeds, RAM etc are important to the end user and that's why do they print them.

As for memory, the bandwidth of each memory controller addressing a chunk of 512MB VRAM is 28Gbps. Through parallel access of multiple controllers and blocks, the effective bandwidth is boosted to a theoretical peak of over 200GBPS. So, its not like one block is slower than the others. In fact they are all capable of same 28Gbps peak. its just that in 970 and other GPUs that use a similar schemes, some blocks cannot be accessed in parallel to the other blocks . No where has nVidia promised that bandwidth would be consistent across all blocks. This was never the case with any GPU released by nVidia or AMD.

I believe the wrong specs sent to reviewers is not a lie because it is a very easy mistake to make and there was no logical reason for nVidia to lie about such a spec. As I said, this sort of thing has happened at both the places that I worked. The marketing teams/documentation/support teams do make mistakes (by sometimes not consulting with the tech teams) and often wrong information gets conveyed. We have had wrong information published in material and caught more than a year after it was published. The core engineers do not always go about verifying every detail posted in the public. Yes, they sometimes turn into PR nightmares. That does not necessarily mean that the company was lying. Companies are run by people and people do make mistakes. I am not saying that companies don't lie for damage control. I have seen the flimsy lies from companies, but in this case, there simply is no logic for nVidia to have lied about ROP count.

The question at the end of the day for consumers is not the architecture that nVidia used, but the performance repercussions. Lets not delude ourselves into believing that we know about GPU architecture better than the GPU engineers. In fact we know less about it than the technical marketing group that sent out the wrong specs to the reviewers. The architectural decisions was the choice of nVidia engineers to make and I don't think they stupid enough to use that sort of scheme in a dozen GPUs if they seriously think it would be an issue.

As I said before, nobody is still sure this partitioning scheme is what is causing drastic performance degradation for certain users.​
 
Lord Nemesis said:
My opinion is that the performance problem is somewhere else and nVidia is yet to locate the root cause themselves or yet to reveal it to public.
Click to expand...
^I bet this.

Look at this way, hypothetical worst case scenario: the specs were faked deliberately. now, knowing the actual specs, would you change your decision and get something other than a gtx 970?
Performance issues are not universal and this is still THE card to have right now.
As far as lawsuits are concerned, none will go far. nvidia can just offer a free game like RD274 says or some refund/discount coupon for their next line of cards.
It really does seem like a mountain being made of a molehill.

On another note, w.r.t to the new internal naming scheme what psyph3r says may be true.
AMD is coming out with HBM this year but nvidia is a year behind so they have to rely on maxwell to counter the radeon 300 series so it make sense for them to wait before rolling out their best chips.
 
^^ Regarding the naming scheme, I would be inclined to believe that it is in sync with the 6xx series.

1st Gen Kepler GTX 680 - GK104
2nd Gen Kepler GTX780 - GK110

1st Gen Maxwell GTX980 - GM204

So it follows that

2nd Gen Maxwell GTX1080 - GM210

Still its not something set in stone. nVidia might release it as a GTX990 or GTX980Ti or something.

http://www.guru3d.com/news-story/nvidia-gm200gm210-gpu-specs-surface.html
http://news.softpedia.com/news/Coun...nch-of-NVIDIA-GeForce-GTX-980-Ti-459689.shtml
 
Last edited:
  • Like
Reactions: Crapmypants
It's pretty obvious this generation though, they've been selling lower tiered products a notch higher than where it's supposed to be, just look at the specs of the 960 and tell me it isn't a lower mid-range card.
All this fiasco is a result of the modifications they made to widen the performance gap between the 970 and 980 which is still not worth the $250 difference.
I'm still not giving them the benefit of the doubt, they're arrogant and have done shady things in the past so it's just nice to see their backs up against the wall for a bit.
 
  • Like
Reactions: Crapmypants
At the moment, the 970 does look to be the card to buy (I plan to buy it right before The Witcher 3 comes out). And looking at the various benchmarks, I don't think the performance hits at the moment can be attributed to this memory segmentation. But it is a concern when we can clearly see a steady increase in VRAM usage by games and the random frametimes spikes once a game starts using >3.5 GB on the 970.

http://www.pcgameshardware.de/Geforce-GTX-970-Grafikkarte-259503/Specials/zu-wenig-VRAM-1149056/

But I also blame a lot of the reviewers. CUDA queries were giving out the correct L2 Cache size and ROP numbers but no-one seems to have caught that. GPU-Z was giving an incorrect number, though I don't know if they query the firmware/drivers for that information or they use an internal database to populate it.

As for NVIDIA and ATI, I won't ever give them the benefit of the doubt. They have done so many shady things over the years, I just see it as a behaviour pattern of "lets-see-what-we-can-get-away-with".
 
  • Like
Reactions: Crapmypants
psyph3r said:
It's pretty obvious this generation though, they've been selling lower tiered products a notch higher than where it's supposed to be, just look at the specs of the 960 and tell me it isn't a lower mid-range card.
Click to expand...
^No argument there. These specs look like a 750ti replacement.
although they might do what they did with the 660 i.e. launch with a 128-bit bus and then put out a ti/boost version with a 192-bit bus.

However, i don't think this applies to the 970. It has a healthy performance lead over the 280x and is priced accordingly.
 
^Yea it is but the entire range is being sold a notch higher just because there's nothing from AMD atm, the 980 should have been a 970, the 970 a 960Ti, a more cut-down version of the GM204 was to be the 960 instead of the rip-off GM206 which is basically the 950Ti.
I don't know how many of these are fanboys or trolls but they got a petition going already, lol
http://www.change.org/p/nvidia-refund-for-gtx-970
 
Last edited:
  • Like
Reactions: RD274 and Crapmypants
That petition is hilarious - it's like the world's thinnest argument for a refund. Shouldn't they be more impressed with how it's doing more with less?

This all has to do with HBM memory. Both amd and nvidia started development of stacked memory however nvidia's solution was delayed or not that good (not sure) and consequently nVidia will be using the kind developed by amd. This gives team red a huge advantage both in terms of performance and getting their products to market sooner.
What i think has happened is both have probably agreed to take it easy on each other for 2014, hence the rebadging of the hd 7000 series and nvidia's weird product placement strategy.
This has probably allowed amd to save money on new gpu launches and given nvidia some wiggle room for their product placement and pricing since they are now going to be a year behind amd.
2015 will see amd taking the performance crown and charging a premium on the pirate island line. nvidia will most likely be forced to compete on price.
 
  • Like
Reactions: psyph3r
Yea they do have an advantage this year but I hope the cards are priced reasonably here otherwise i'll have to settle for something like the 370x (most probably a rebadge?) for the time being, this is the first time ever that my system is running w/o a gpu since i got rid of my second 6950 last month and i'm not liking it.
Nvidia delayed their implementation of stacked memory due to problems at TSMC iirc, not sure thg.
 
AMD Cashes in On GTX 970 Drama, Cuts R9 290X Price.

150a.jpg


AMD decided to cash-in on the GeForce GTX 970 memory controversy, with a bold move and a cheap (albeit accurate) shot. The company is making its add-in board (AIB) partners lower pricing of its Radeon R9 290X graphics card, which offers comparable levels of performance to the GTX 970, down to as low as US $299.

And then there's a gentle reminder from AMD to graphics card buyers with $300-ish in their pockets. With AMD, "4 GB means 4 GB." AMD also emphasizes that the R9 290 and R9 290X can fill their 4 GB video memory to the last bit, and feature a 512-bit wide memory interface, which churns up 320 GB/s of memory bandwidth at reference clocks, something the GTX 970 can't achieve, even with its fancy texture compression mojo.

http://www.techpowerup.com/209412/amd-cashes-in-on-gtx-970-drama-cuts-r9-290x-price.html
 
Status
Not open for further replies.
Top Bottom