The Real Nvidia GeForce GTX 970 Specifications

Page 2 - Seeking answers? Join the Tom's Hardware community: where nearly two million members share solutions and discuss the latest tech.
Status
Not open for further replies.
I wonder what reviewers will think of a GTX 960 Ti if it is indeed a 25% cut-down GTX 980. That's a whole GB of VRAM in the gutter, lol.
The argument over the GTX 970 is that the advertised 4GB of VRAM and 256-bit memory bus cannot be utilised to their full potential because of a disabled component in the memory subsystem. If they cut this GPU down for a GTX 960 Ti by disabling 1/4 of the memory subsystem and including 3GB of VRAM, you would be left with a 192-bit bus and 3GB of VRAM that would not have the issue of the GTX 970.
also having big pagefile on SSD also helps it seems,
(they notice the pagefile increased largely after go above vram barrier)
so try limit the pagefile to 2-4 Gb and soon after their system crash
If the system page file is being used above 2GB, they have managed to exceed both the VRAM on the card and the system RAM. It is hardly Nvidia's fault if this causes slow performance, and limiting the page file beyond where it is required is likely to cause a crash in Windows under any circumstances.
Even the last 512MB VRAM on the card is still faster than accessing system RAM through the PCI-E bus, and this is orders of magnitude faster than the page file.
Simply put, the tester is using settings that their system is not capable of running and a GTX 980 or R9 290X would not fix that.
 
It doesn't really matter?? Jeezzzz, exactly what Anand and every other prepaid said, leaving little doubt this is just NV's PR damage control blitz.

The issue here is not that the card will magically start under-performing because of this "discovery", the issue is that this is, plainly put, False Advertising. Consumers were mislead and NV has to be held responsible for that.
 


Agreed, But I don't remember how much ram it has 16 GB or 32GB.. (or it's setup/rig)
I'm too lazy to re-read the 800+ post on OCN forum..
(one of few first forum that rise up this issue)

That's why I'm asking Toms to do investigation about this..

edit: BTW, I'm sure that tester have more than capable rig,
he just "testing" what page file have effect in this matters
 
False advertisement is False advertisement whether the card need full 64 ROP /L2 cache or not.

I hope they honored a full refund to owners that wanted to refund.
 
that's a horrible metaphor, for many reasons. it's still false advertisement. it's unethical and it's not fair for the consumer. it's not fair for the people who purchased the product to not be able to use it to it's full potential. i don't think software can fix this failure and i don't think it's a complete product.
 
In my view; I would love to show up my care with the V8 engine, when I pop up the trunk. I would love to show up my graphics card capability with a see thru case and the high FPS result. I could tell my friends, some items are disabled, but still a fast card, and my friends don't care, as long as visually it looks cool. It will be a concern with highly techie guy though.
 
*This is old news so what's going on?
http://www.guru3d.com/articles-pages/gigabyte-geforce-gtx-970-g1-gaming-review,5.html

It has already been reported that the GTX970 performed slightly worse in some high resolution games than expected in comparison to the GTX980.

We then learned this was due to the slightly lower ROP count as this was used to decode the lossless compression but it couldn't keep up on the GTX970 in some scenarios.

THUS, this isn't news to many people so how is it suddenly news now?

The linked article at top is from September 2014 and it's correct showing 56 ROPS.
 
The 970 bus It is not 256bit.
Its 224bit max (for the 3.5GB), and the last .5GB can only be accessed via a 32 bit controller.
(This is all documented by nvidia in their latest GTX970 graph)
 
Also the 224bit bus and 32bit bus cannot be access simultaneously.
So the third error in your chart is listing the bandwidth at 224GB.

According to nvidia specification, the 970 got 196GB of bandwidth not 224.

And this is verified by all benchmark to date.

Please correct your incorrect chart.
3.5GB on a 224bit bus , with 196GB bus bandwidth.
+ .5GB of 32bit 'sideband' memory with 28GB bus bandwidth (cannot operate in parallel with primary 224bit bus)
 
Whether it was a mistake due to misunderstanding it doesn't say anything good about NVIDIA (not saying they are bad). The GTX 900 series have been released for a while now and they had all that time to correct their mistakes and release their statements before the error was found by someone else. Nothing changes the fact that people bought the GTX 970 thinking it had better technical specs. NVIDIA is lucky that the performance is still great but you can't stop people from thinking that they paid more than what they should have as the GTX 970 technically was not a 64 ROPS and a 2MB L2 cache video card.

The video card comes with 4GB of VRAM but they designed it to not use it if possible because the remainder 512MB of VRAM would perform slower. There is a tiny bit of bottleneck whether it's noticeable or not. If it's designed not to use the remainder .5GB then a user would almost never use the full 4GB making it a 3.5GB GPU. Not that it's a deal breaker.

At the end the GTX 970 is still a monster of a card that is pretty much the best GPU overall for the money in power and efficiency. Two could be had for a little over a GTX 980 and perform better while it performs pretty much better than two GTX 960s that cost more than a single GTX 970.
 


Now I'm confused..

From I know guru3D (and some other) have to change their ROPS count in the review after NVIDIA released statement that said old ROPS count is wrong..

since I'm not their member or familiar about how they write/edit on article then I'll let other confirm it..

Btw this issue is since day 1 issue, most people don't see the problem,

some see it but don't have explanation for it so they overlooked..
(some using card for CUDA rendering and get vram allocation error, I read somewhere)

some see problem keep probing it, make a big fuss, and gain momentum..
and lead us here....
 
rdc85,
The ROP count news can't be recent since I read about the ROP count difference months ago.

Also the "segmented" memory as described makes absolutely no sense at all. It's the same 4GB of VRAM physically so what possible benefit would there be to splitting the memory? None.

The ROP units help decode the data which gets compressed before going across the video bus (narrower bus is cost saving). People complained about bus bandwidth for the GTX970 and GTX960 without understanding this.

Long story short it's a non-issue since performance is as advertised.
 
"segmented"
I wish I could EDIT. I meant to say the article could be confusing, not that the memory wasn't being reported in software.

If you read carefully it does state no real-world difference and that all 4GB is accessible.

I am a little baffled though how this is "reported" since I can think of no logical reason why the card would give an OS the details of this supposed segmentation since there's no reason an OS would need to know that I can think of.

It would be like telling Windows it was GREEN in color.
 

It is more than just the ROP units that are disabled, it is also the direct interface from one of the memory controllers to the crossbar:
http://www.pcper.com/reviews/Graphics-Cards/NVIDIA-Discloses-Full-Memory-Structure-and-Limitations-GTX-970
This is what makes access to the last 0.5GB slower, and why they segmented the memory into a primary pool and a secondary pool.
 
I have a friend who bought 2 of these to run in SLI with a brand new 4k monitor he got and when playing his favorite game; Shadow of Mordor, at 4k with all bells and whistles on, he was having crazy performance drops, at first he thought it was from his ssd, then from his overclock on the cpu, then he thought it might be the ram, then he lowered aa and things got better, now we know what was going on.
 
For people asking about the 3.5 vs 4.0 GB issue, the card is indeed 4.0GB and does use all 4GB. The drivers are the ones doing the memory management, they have a cacheing algorithm that tries to keep frequently used data in the 3.5GB and only puts data in the last 512MB segment if the first is full and even then it only moves old data into that slow segment while putting new data into the fast segment. This is what started the whole thing to begin with, someone noticed their card didn't want to go above 3.5GB of graphics memory usage and decided to try to force it to. That last 512MB was being used, but as victim cache for data to go before it's removed from the fast 3.5GB memory range. In practical application the only way users would naturally run into a situation where they are using more then 3.5 but less then 4.0 GB of graphics data is to run at settings that would likely be a slideshow for this card. The 970 doesn't do 16x AA 4K Ultra settings, and contrary to what most people think it's the AA that use's all the memory. You can run 1440P no problem, even 4K, provided you don't try to max out the AA setting (AA cause's a 2~4X increase in memory texture sizes, sometimes more).

I do agree with the author that inventing conspiracy theories to explain plain human greed and stupidity isn't very productive. Working for a large multi-billion USD global corporation I've seen similar communication failures. Tech guy A gives report to marketing guy B (who is definitely not tech savvy). Marketing guy then does the initial release while paraphrasing tech guys words. Somewhere in that process marketing guy writes something that's not true and nobody catch's it before the big brief to the boss and subsequent green light for press release. After press release, or even the big boss green light meeting someone noticed it and brings it to a middle manager who then has an hushed emergency meeting where they discuss the mistake. The most important question is going to be "what is the chance of someone finding out before our next product release", and there is going to a guesstimate of "pretty damn low". Middle manager then decides to keep the entire thing hushed and wait out until next major product cycle. Senior executives are rarely informed.

We make jokes about Office Space type situations, but they do happen and are more frequent then most would think.
 


Segmented, actually partitioned NVIDIA exact word...

3.5Gb in high priority high speed partition and 512mb in lower priority partition..

why they chose that beats me...

It will confusing if not following the story from start
which i happen to follow at mid Jan, not post any since no concrete proof or i own 970 just forum browsing.
Now it's out in open, NVIDIA has admitted it. I sharing what i have read and opinion i have here...

anyways i recommend go to OCN forum or read PC perspective article for more info..
NVIDIA forum also useful but i never looked there...
(and i heard the mod there has been doing forum cleaning lately)

edit: from what i know..

- 970 is cut down from 980 so it carries some limitation..
apparently the memory management/portion also infected.. something nvidia PR said it doesn't at first..
- To counter this weakness/limitation Nvidia relay on driver to keep last partition not accessed
until it's absolute needed to make sure smooth game play experience
- some will have issue (high end system, SLI, CAD, 4k, etc) but mostly don't,
since the driver did a good job or their system don't push the card to limit hence 970 is still a good card..
- some raging for the issue, some feels cheated their 4 Gb card turns out 3.5 Gb card,
and some troll fishing
- also speedy big page file helps (above 20gb or let system managed)

I just demands more testing/benchmark 😀
(and it's been a while we don't get interesting thing to discus/read/comment on haha)
 
what im most curious about is what alternative card these customers where going to buy instead of the 970 for $330. the 290 took almost a month after the 970s release to drop to about $300(after rebates). around thanksgiving/black friday(over two months after 970) it dropped to around $270 and it has stayed there since, besides the random deals you might see around $250. im very confused about all this buyers remorse i guess, it just doesn't seem logical.
 
the only thing can beat this card is 980 or 290x (if u don't mind the power/heat) from i know.. (for sli / 4k)
hardware chuknuk have "preliminary" benchmark about this memory issue..

don't know they reliable or not. i just hope tom's will do the same..
 
Can a new bios fix the problem? If so, cool. If not, Nvidia needs to do something for the people who bought these GPU's.

I would be cool with a $100 Steam Wallet credit :)
 
I'm super smart, so I make my processor decisions based on my own projections of what various chips' specifications will mean for my performance needs, and NOT based on the actual benchmarks that show the actual performance. That means a company like nVidia can EASILY take advantage of people like me, since we don't read benchmarks from sites like Tom's but rather just read esoteric technobabble specification numbers!

AS IF NVIDIA DIDN'T KNOW THAT'S HOW WE SHOP!!! Liars and thieves and cheats. I hope they all go to jail.
 


Doubt it, since it not a flaw. it design weakness result from cut down...

It working as intended..
this bring me hard to believe it just "miss communicating" from PR department
 
The author needs to look into this situation a bit closer, the facts have been out there for a week.

1) the last 500mb of vram works at ddr3 speeds, and can't be used in conjunction with the rest of the ram

2) the card's drivers automatically detects benching software and intentionally prevents the last 500mb of vram from being used by the benches; this is why the issue wasn't apparent, and doesn't seem to affect anything. Due to the drivers being coded to hide this

3) the drivers will actually prevent the cards from using the last 500mb of vram unless absolutely necessary. as a result you'll get games which will hit 4gb of vram usage on the 980, but won't break 3.5gb of vram usage at the same settings on the 970

4) when the card HAS to use the last 4gb of vram it starts to stutter badly.... far worse then the worst xfire setups prior to amd's work on fixing xfire stuttering. No "standard" benching games will hit 4gb of vram usage at full settings in less then 4k... and as you know, nvidia's fcat tool doesn't work on 4k resolutions so frame pacing benches which would have exposed this issue never showed up when reviewers were reviewing the card.

I hope the author takes this information into account when he digs deeper into this issue in the future.
 
Status
Not open for further replies.