UPDATE 1/29/15: This forum post has since been edited and basically removed, with statements made on Twitter that no driver changes are planned that will specifically target the performance of the GeForce GTX 970.
The story around the GeForce GTX 970 and its confusing and shifting memory architecture continues to update. On a post in the official GeForce.com forums (on page 160 of 184!), moderator and NVIDIA employee PeterS claims that the company is working on a driver to help improve performance concerns and will also be willing to "help out" for users that honestly want to return the product they already purchased. Here is the quote:
Hey,
First, I want you to know that I'm not just a mod, I work for NVIDIA in Santa Clara.
I totally get why so many people are upset. We messed up some of the stats on the reviewer kit and we didn't properly explain the memory architecture. I realize a lot of you guys rely on product reviews to make purchase decisions and we let you down.
It sucks because we're really proud of this thing. The GTX970 is an amazing card and I genuinely believe it's the best card for the money that you can buy. We're working on a driver update that will tune what's allocated where in memory to further improve performance.
Having said that, I understand that this whole experience might have turned you off to the card. If you don't want the card anymore you should return it and get a refund or exchange. If you have any problems getting that done, let me know and I'll do my best to help.
–Peter
This makes things a bit more interesting – based on my conversations with NVIDIA about the GTX 970 since this news broke, it was stated that the operating system had a much stronger role in the allocation of memory from a game's request than the driver. Based on the above statement though, NVIDIA seems to think it can at least improve on the current level of performance and tune things to help alleviate any potential bottlenecks that might exist simply in software.
As far as the return goes, PeterS at least offers to help this one forum user but I would assume the gesture would be available for anyone that has the same level of concern for the product. Again, as I stated in my detailed breakdown of the GTX 970 memory issue on Monday, I don't believe that users need to go that route – the GeForce GTX 970 is still a fantastic performing card in nearly all cases except (maybe) a tiny fraction where that last 500MB of frame buffer might come into play. I am working on another short piece going up today that details my experiences with the GTX 970 running up on those boundaries.
Part 1: NVIDIA Responds to GTX 970 3.5GB Memory Issue
Part 2: NVIDIA Discloses Full Memory Structure and Limitations of GTX 970
NVIDIA is trying to be proactive now, that much we can say. It seems that the company understands its mistake – not in the memory pooling decision but in the lack of clarity it offered to reviewers and consumers upon the product's launch.
A big issue with this whole
A big issue with this whole deal is how this inudorinately affects SLI users at something like 4K or 1440p. Where they have enough shading power to push the options and AA, but the card will probably buckle and stutter under the weird RAM set up.
I guess we’ll see what
I guess we’ll see what happens after the driver update.
Apparently there is NOT a
Apparently there is NOT a special driver coming from nvidia. PeterS@nvidia redacted his statement:
“Hello,
I’m sorry that what I wrote was poorly worded, I realize I made it sound like there was a special patch or something and that is definitely not the case. We are always working on new drivers that tune performance and add features, the GTX 970 is no different.
Are you having a specific issue with the 970 that I can help you with? Unfortunately we made an error in the reviewer guide but the GTX 970 is one of the best GPU’s we’ve ever built.
http://www.tomshardware.com/news/nvidia-geforce-gtx-970-specifications,28464.html
Let me know how I can help.
”
From:
https://forums.geforce.com/default/topic/803518/geforce-900-series/gtx-970-3-5gb-vram-issue/219/
What will “probably” happen
What will “probably” happen is that review sites will test the heck out of this and will find that the worst-case scenario produces a negligible frame rate drop.
This is a textbook case of people that don’t really understand what’s going on turning the mole-hill they are looking at into a mountain in their minds.
The customer is always right,
The customer is always right, and they should return the cards, why wait for an update to “Fix” the problem. RMA the cards, and get something that does not have the gimped memory channel. This problem is going to put a big asterisk by these cards, and Nvidia, and its retailer pipeline are going to have to sell them at cost, just to recoup their expenses. Next time do not overload the memory channel, whatever, and try a gimped solution.
I agree that they should take
I agree that they should take them back if people really want to return them but this “issue” really effects very few users. People do not seem to understand that almost all chips sold will have a defect in them somewhere. It is not that uncommon to have things like a few extra cache blocks to make up for defective areas without needing to disable the whole cache. The clock speed is obviously binned for; not all can run at the same speed. Often an entire product line-up is made from the same die just with different portions disabled. Currently, it loos like the 970, 980, and all three mobile parts are actually the same die with different areas disabled.
Presumably people bought the 970 because they looked at the performance in the games they played and it was the best value. If they want to return it and get an AMD part that may perform worse, but has a full 256-bit buss then that is their decision.
They could also pay the extra ~220 dollars for a full 980 part.
Binned dies with defective
Binned dies with defective units, plus some extra fusinng to make the binned parts perform equally, and one memory channel sharing double the responsibility/load in high resolution situations. Nvidia has made a big mistake in the eyes of the customer, and we all know that games as well as benchmarks can be cherry-picked in reviews. There are no better reviews than the actual customers having problems, with their respective games, and system builds, and any of these lower grade derived SKUs are almost always the result of binning, and users need to wait a while before purchasing any GPU parts that they suspect are the results of binning, and let the first buyers be the guinea pigs.
Return the parts, and wait for something better, these parts will be on fire sale, because of the big “*” buy their name, and AMD is about ready for some new(non Binned) releases. No gimping on channels please, its not the same as fusing of some SPs, etc.
Can’t wait for a third player in the discrete GPU business, and Imagination Technologies would probably not say no, if one of their customers ordered a discrete part to be made. Those PowerVR GPUs are getting more impressive with each new generation.
All chip makers bin or
All chip makers bin or salvage different dies for different products. Intel’s highest end desktop parts, the extreme edition core i7s are actually salvaged Xeon server parts with defective cache or cores. There is nothing wrong with binned or salvaged parts. It is a standard practice. The only issue is that they did not communicate correct specifications.
I don’t know where PowerVR GPUs stand compared to current desktop parts, but they do not seem to be of too much interest. There should be a big jump in performance soon. Performance has been somewhat stagnant for a while due to being stuck at 28 nm for several years. In the next generation or 2 we should get a smaller process node plus stacked HBM memory. This should be good for a large jump in performance. The stacked memory could be somewhat disruptive since it would allow a CPU sized device to reach dedicated graphics levels of performance.
All chip makers bin or
All chip makers bin or salvage different dies for different products. Intel’s highest end desktop parts, the extreme edition core i7s are actually salvaged Xeon server parts with defective cache or cores. There is nothing wrong with binned or salvaged parts. It is a standard practice. The only issue is that they did not communicate correct specifications.
I don’t know where PowerVR GPUs stand compared to current desktop parts, but they do not seem to be of too much interest. There should be a big jump in performance soon. Performance has been somewhat stagnant for a while due to being stuck at 28 nm for several years. In the next generation or 2 we should get a smaller process node plus stacked HBM memory. This should be good for a large jump in performance. The stacked memory could be somewhat disruptive since it would allow a CPU sized device to reach dedicated graphics levels of performance.
I used to use GTX 680 SLI
I used to use GTX 680 SLI (2GB cards)in my gaming rig at 1440p without any problems. Most games aren’t that demanding unless you really crank up the AA or you’re running something like Crysis 3 at 4K.
I’ve since moved it to the HTPC which is hooked up to a 4K TV, and games like Tomb Raider run fine on it.
Guru3d have done some testing
Guru3d have done some testing with Shadow of Mordor at 4K and passing the 3.5Gb mark, and have found… not much impact beyond the expected general drop in performance. Whatever, if anything, the driver is doing seems to be working pretty well, at least for them.
It’s good to see that the
It’s good to see that the performance difference does indeed seem to be minimal as Nvidia stated. This upcoming driver should help, too.
With all that said Nvidia still screwed up. It’s not a huge deal, but I still think that someone from high up in Nvidia should just come out and apologise and admit they should have fully disclosed the 970’s memory architecture from the beginning. Maybe a free game code or something for 970 owners as a sign of good faith would be good, too?
A German tech website has
A German tech website has tested the GTX 970 vs a GTX 980 clocked down to 958 Mhz – which would be the same performance as the GTX 970 (in sense of TFlops afaik). In Watch_Dogs at UHD with ultra settings the GTX 980 (which does have the same memory/cache/ROP setup as originally advertised by Nvidia) was a whopping 34% faster. The GTX 970 also started stuttering.
The GTX 980 has 384 more
The GTX 980 has 384 more cores than the GTX 970… Putting them at the same clock speed does NOT equate to putting them at the same performance.
Not same clock speed,
Not same clock speed, lower…
958 MHz vs 1178 MHz.
It’s not that easy to make
It’s not that easy to make ‘equivalent speed’ that way. If the bottleneck is something easily parallelisable but requiring minimal actual computing time (e.g. something ‘gated’ by a minimally-parallelisable step), then the slower but wider card will perform faster. Likewise, if something is not easily parallelisable then the faster but narrower card will perform faster.
Graphic rendering is a mix of both (pixel tasks are often embarrassingly parallel, but vertex operation often are not), so just fiddling with clock speeds between two different cores will not make them perform in the same way.
I can only assume that they
I can only assume that they choose exactly these clock speeds, because they offered equivalent performance before the 3.5 Gigabyte VRAM partition was exceeded. At least the article states that the GTX 980 was clocked to deliver equivalent performance as the GTX 970.
Geek3d has done some testing
Geek3d has done some testing with the OpenGL side of things. An you’ll be shocked about what they’ve found out.
http://www.geeks3d.com/20150127/opengl-apps-not-impacted-by-gtx-970-vram-limitations/
The conclusion.
According to
The conclusion.
Is windows under directX just
Is windows under directX just not using the second partition?
this needs to be tested
this needs to be tested thoroughly using FCAT
also Nvidia should launch a 970 V2 with the full memory/ROP/L2 as the original specs they released for the original 970.
Physically impossible. The
Physically impossible. The way maxwell architecture is setup when those cores are disabled the memory controller and cache goes with it. Can’t separate it.
As expected, PCper are the
As expected, PCper are the first guy’s to post about nVidia and their damage control and at the same time reiterating how they don’t think what they did is a big deal.
Looking forward to tonight’s podcast where I’m sure it’ll be the guy’s saying how the internet overacted and everything was blown out of proportion.
It kind of did, and we likely
It kind of did, and we likely will.
But you know, it's comforting to know that every other media outlet at least shares our opinion.
That has to mean SOMETHING, right? We can't ALL be paid-off, "BIOS-ed" shills, right?
there might be overreaction
there might be overreaction regarding the performance implications, BUT what nvidia did is very serious, intentionally or not (I think it’s was not intentional) they lied about the specs (8ROPs and 256KB l2 are simply not there), and you could say they lied about the memory amount and bandwidth, because I don’t see this card using the full 256bits while gaming, just a pool with 224 and another with only 32bit, it’s simply not a 4GB card in the same sense as the 980 or 290, or a 256bit card like the 980… it’s a 224 3.5GB card with an extra 512MB 32bit “buffer”,
what is funny is that Nvidia only messed up with the 970, all the other 4 or so GM204 based cards have a clean memory/ROP/l2 setup (each memory controller active have a ROP/l2 available, unlike the 970)
Yeah, so you just have to
Yeah, so you just have to believe that not a single engineer or marketing person from nVidia saw a review out there that stated the specs wrong. How long has the 970 been out? Not buying it.
Thank you to the community however for finding out the BS that nVidia clearly didn’t want to share and the hardware sites couldn’t figure out.
Absolutely agreed.
It is
Absolutely agreed.
It is conceivable (though still unlikely for such a large and well-funded marketing department) that incorrect information would be mistakenly given to reviewers on launch, but it’s laughable for Nvidia to imply that for all the months since launch, not a single engineer has seen any review or discussion of the 970 that mentions incorrect specifications. It’s true that the 970 is still not a terrible card, but this is textbook false advertising that Nvidia clearly hoped they could get away with, even in the remote possibility that it was originally an honest mistake.
“I don’t believe that users need to go that route [returns]” -Ryan
It’s almost disgusting to see how eagerly the press scrambles to apologize for Nvidia’s blatant bullshit. Assuming you aren’t simply being paid for these comments, you are encouraging this kind of absolutely anti-consumer behaviour from Nvidia (wasn’t the point of doing reviews to protect the consumer?). If I had a 970, I would attempt to return it no matter how satisfied I was with the performance – Nvidia must be punished for this false advertising, because if they are not, they will try it again. It’s not so hard to imagine that if it was known that the 970 had less (“usable”) memory than the comparably priced R9 290(X), at least a few people would have been swayed towards AMD.
Actually, several sites have
Actually, several sites have called BS on NVidia’s excuse of misinformation.
You and Allyn truly deserve
You and Allyn truly deserve medals for being so patient with some of the people leaving comments here. ;-0
We do it because every once
We do it because every once in a while a post like yours brings a smile to our face and makes it all worth it. Thank you.
Think how much better it
Think how much better it would be if most posts where positive. Just a thought :p Say hello to Jen.
Let it be said, to never
Let it be said, to never completely trust any website that gets advertising revenue from the same industries/companies whose products are reviewed. There will always be questions related to objectivity. PCper is fairly unbiased, but Tom’s, and now AnandTech(not owned by Anand anymore), well, there goes the technology reporting industry. If you want more unbiased technical reviews you are going to have to go behind the pay-wall.
The entire Technology industry is so full of hucksters, shills, and stock manipulators(for those that are shorting/whatever stocks), and do not forget the large Chip/OS monopolies(every bit as bad as the Standard oil trust, and Ma Bell, of the past) all trying to hold onto their market share, by belittling the competition, and arm-twisting the device OEMs.
Technology is the new moneymaker of the past 3 decades, expect there to be much snake oil, and underhanded trickery, nothing has changed much from the gilded age!
–See the Glory, Of the Royal Scam…
Fully agree. It must really
Fully agree. It must really suck at times being a reviewer, since you have to maintain a calm and professional posture under all circumstances. It’s always the frantic few that act like they’ve been kicked in the balls.
I installed my GTX970 Phantom into my NZXT Rogue HTPC/Gaming case yesterday, fired up Far Cry 4 and went for Ultra. What I got was Ultra performance, Ultra low noise and an Ultra big grin on my face. 2015 is good. 🙂
OK, you know the very very
OK, you know the very very bad answer to this. It’s very rude so I apologize in front. But when we are dealing with Nvidia someone could also say that maybe the press also shares the same bank account. There. I told you that it was bad and rude. I apologize again.
Anyway, when AMD’s Hawaii chip was throttling and the speed was not 1GHz, most of the tech sites, especially a few green ones (-Hello Tom!), where looking at the GPU frequency and saying that “Who gives a s#$%^t about performance? AMD lies about the GPU’s frequency”. Now, everyone is covering Nvidia’s lies by downplaying the problem and saying that it is no big deal. It’s only -3% difference. A totally different approach when in fact what Nvidia did was much worst than what AMD did.
This is a classic example of
This is a classic example of how easily things can spin out of control on the net and how irrational people can be. On Sunday morning I woke up to read a newsflash on tweakers.net stating that a ‘memory issue’ had been found with the GTX970. I had ordered the card the previous evening, so I wanted to get more info on the matter ASAP, which I found here.
A big thumbs up to you guys for taking the matter seriously and putting things into perspective. Keep up the good work.
The way I understand it
The way I understand it ,lower performance. Is only when :only the 512 mb would be needed (example :chess game)I don’t think any would care lol.all other case should run at max,hell this might actually be good .why?MSI/msix can be done on the 512 MB part instead of on the ram (GPU part)
lower performance yes, the
lower performance yes, the old reviews are accurate in terms of performance,
but in terms of the fake specs the effect is always there, since they promised a 64ROP/2MB l2 256bit card and delivered a 56ROPs/1.75MB l2 224bit card
Some more
Some more here:-
http://www.hardwarecanucks.com/reviews/video_cards/the-gtx-970s-memory-explained-tested/
Now that the cat is out of
Now that the cat is out of the bag regarding
the specifications, does the GTX970 have
256-bit Memory Bus Width or not?
physically yes, but in
physically yes, but in practice it’s a 224bits card at best, since the other 32bits are not aggregated to the total, but just an additional, separate 32bit bus space.
Ryan ,i am still waiting for
Ryan ,i am still waiting for ur update
Ryan, can you get a response
Ryan, can you get a response from Nvidia on how the GTX660ti memory system works like the one they gave about the GTX970?
Anandtech already mentioned that is very similar.
But details would be nice.
660 Ti has three 64 bit
660 Ti has three 64 bit memory controllers (192bit total), two connect to 512MB of memory, and the third connects to 1GB of memory. At least that’s how I understand it. It’s still a single pool of memory, and not comparable to the unique configuration of the 970.
Anandtech says something
Anandtech says something else:
“The end result as it turns out is very similar, and while NVIDIA has never explained in-depth how they handle memory allocation on those cards, it turns out that it’s very similar to GTX 970’s memory segmentation. Which is to say that NVIDIA actually has multiple generations of experience with segmented memory, and this is not the first time they have implemented it. Rather this is first time we’ve seen such a configuration on a high-performance card such as the GTX 970.”
http://www.anandtech.com/show/8935/geforce-gtx-970-correcting-the-specs-exploring-memory-allocation/2
I do stand corrected. Thanks.
I do stand corrected. Thanks. That definitely made for an interesting read.
I don’t agree with
I don't agree with Anandtech's position on that. In the case of that card, there were larger proportional chunks with less of a bandwidth delta between them. In this case you have a 7:1 speed ratio between the two segments of memory. Calling it 'very similar' is a bit off.
Thanks a lot for your view on
Thanks a lot for your view on that!
I got curious after Nai’s test in headless mode have given this results:
http://i59.tinypic.com/29w4bd1.jpg
http://oi59.tinypic.com/29w4b
http://oi59.tinypic.com/29w4bd1.jpg
Allyn is correct. AnandTech
Allyn is correct. AnandTech is wrong about the 970 bus configuration being similar to the 660Ti. They are totally different.
Fcat would have been produced
Fcat would have been produced ALREADY if this were an AMD card, yet the creator of the software to test AMDS XF MS is somehow gimped now, and can’t be used to show the performance of it’s own hardware.
Not until the FCAT code is adjusted, or very creative drivers are coded to hide the inefficiency of the last 500 anyway…
Those results could have been produced by Saturday night without issue, Monday night at the very least if folks didn’t think it was “Important enough”…
It’s Wednesday afternoon, and we are reading PR about how it was a mistake….still, and with folks saying it’s perfectly fine the way it is.
Sure made a lot of $ on these cards, I’m certain someone over there could have produced FCAT by now, especially since the entire weekend was a fire drill with conference calls and meetings….was not a good weekend for some over there, I’m betting on that….
Still no FCAT…..pathetic!
Did you bother to read the
Did you bother to read the news post?
“I am working on another short piece going up today that details my experiences with the GTX 970 running up on those boundaries.”
There was no update to FCAT
There was no update to FCAT to handle the 900 series cards, nor are we using an updated version of it for our retests now. Same driver, too. How's your argument hold up given that information?
FCAT takes time, regardless of GPU being tested, especially in a case like this where we are trying to determine a very small performance difference, and then have to perform the same exact round of testing on another card that does not have this same segmentation. Then we have to try and put that data into graphs that properly show the performance delta from *this* issue as opposed to the performance delta resulting from the fact that one card is overall slower than the other card in the first place (for reasons aside from the segmentation).
Seems some sites have actual
Seems some sites have actual nV employees posting on the forums, peter at GF.com, Brian at TR.com…
SAME answer/post given on both sites, the post is a copy/paste on both sites, so this was clearly very carefully thought out.
Does mention that if you need help getting a refund to contact he/them directly.
it’s a social media guy from
it’s a social media guy from NV, he is posting in many different websites, nothing wrong with that
Seems they have people on
Seems they have people on pretty much all sites with some traffic. A swedish forum I read have an official rep now. Showed up today (28 jan) and have made 10 posts so far.
Everything is OK, nothing to see here. Circulate, circulate.
Peter said he is an Engineer
Peter said he is an Engineer
Now we know who the
Now we know who the sacrificial lamb is.
FCAT?
FCAT?
http://treasure.diylol.com/up
http://treasure.diylol.com/uploads/post/image/514750/resized_winter-is-coming-meme-generator-brace-yourselves-fcat-is-coming-0ad6b2.jpg
^ +1
^ +1
eheheh
Just DON’T FORGET SLi!
eheheh
Just DON’T FORGET SLi!
FCAT = Flat Cat!
FCAT = Flat Cat! https://www.dropbox.com/s/8f4qzewch91ylwo/004.JPG?dl=0
Placebo pills, placebo
Placebo pills, placebo (wonder) drivers for the customers.
Such as the OMEGAs, you mean.
Such as the OMEGAs, you mean. Because, I had 0% observable impact in performance going from 14.11 to OMEGA on my AMD rig.
AMD Omega drivers have no
AMD Omega drivers have no impact on Nvidia hardware, so it is logical for you to have 0% impact on performance.
high resolution textures,
high resolution textures, 1440p, and 4k will murder this card in single and sli setups plain and simple. Anything that has to address that last 512mb of memory will hit a performance WALL. PCper should hold themselves to a higher standard than being nvidia shill.
224 GB/s aggregate
196 GB/s (3.5 GB)
28 GB/s (512MB)
memory bandwidth numbers taken from here
http://www.tomshardware.com/news/nvidia-geforce-gtx-970-specifications,28464.html
yesm the 196GB/s is the best
yesm the 196GB/s is the best it can achieve, you can’t add the other 28GB/s to it.
224 is what the 980 can achieve, definitely not the 970, 970 is 196 for 3.5 and 28 for 0.5; 980 is 224 for all 4.
just to add, mentioning 224
just to add, mentioning 224 for 970 is very misleading.
Tea & Sympathy
“I can’t
Tea & Sympathy
“I can’t guarantee anything specific like refund vs credit vs upgrade+cash or whatever as they are independent companies.”
“But since we don’t actually sell cards directly (aside from a limited run at Best Buy), I can’t force our partners to do vouchers or anything like that.”
“To be extra clear, I think the GTX970 is a great card and honestly don’t believe there is a better card for the price. If you are unhappy for whatever reason, talk to your board mfgr and give them a chance to help you out. If that route fails, let me know and I’ll try to help.”
Tom Petersen for all the
Tom Petersen for all the hoopla and technical expertise and close relationship with his engineering team still didn’t know about the spec differences?
You guys asked him about it at GTX 960 launch off-air.
I guess that means he is still just a marketing guy like his title implies.
Every time you have him on its one big Nvidia commercial.
“Every time you have him on
“Every time you have him on its one big Nvidia commercial.”
Honestly there’s not much else to expect when you get an industry guest on the show – a company isn’t going to pay an employee to visit and talk unless they think it will somehow improve sales. There’s nothing inherently wrong with that, but it is interesting to see how differently Ryan treats Tom and Richard Huddy, for example. Huddy gets all the difficult, aggressive, and interesting questions (as it should be for both sides), while with Tom, Ryan (as one commenter eloquently put it) just blows sunshine up his ass.
That bias treatment is all
That bias treatment is all too obvious. I suspect that’s why some view PCPerspective as a shill.
I have to say, that Josh gives them a much more balance treatment then the others. It could be his age or just that he is tipsy still try’n to figure that one out.
Josh is the best pcper
Josh is the best pcper writer: fair, knowledgeable, and lewd.
I typically learn something
I typically learn something new whenever Tom is on. He's a smart and personable guy who is obviously passionate about what he does. Same with guys like JJ. They are representing their companies, they will say lots of nice things about their products, but they will also take a deep dive into what they are doing to help showcase these products. Knowledge is power! Just make sure to read around and gather as many differing opinions as possible and weigh it all together. We have brains, be sure to use them!
In the end, this will all
In the end, this will all blow over I bet. This whole thing will be like Shakespeare’s classic play “Much Ado About Nothing”. 😉
Even with 3.5GB and 196GB/s
Even with 3.5GB and 196GB/s it manages to perform pretty well. I suspect thatsall the bandwidth saving they did with the compression.
It seems like Nvidia is guilty of trying to make their specs look better than they really are and consumers are rightfully pissed.
This is from an “Nvidia fanboy” who still says that Crossfire, and AMDs CPUs and GPUs architectures suck. Ive got a 780 and im glad now lol.
Now if only people would get this angry about things like “LED monitors”, light bulbs, food, cars, and all the other stuff theyre constantly lied to about.
Im not saying its not irritating to be treated like idiots, and it looks like Nvidia really did cheat here to make the 970 look better on paper than it actually is.
It also seems like a waste of resources to design the memory subsystem to keep a useless portion of .5GB(1/8) of the total amount especially if its just to make it technically 256bit. Its almost definitely nowhere near 224GB/s in real world usage.
Im just saying: go get this pissed off over the other things i mentioned. Theyre really much worse lol.