WCCftech: Memory allocation problem with GTX 970 [UPDATE] PCPer: NVidia response

[ocre]

There s only 208 bits active out of 256 when it comes to feeding the SMMs, 48 bits are inactive computing wise and thoses ones are used for the datas of the 768MB pool, do the maths starting from 4096MB, you ll have 3328MB adressed with the first 208 bits and 768MB adressed by the remaining 48 bits.

The discretanpcy comes from the fact that the 768MB adressed by the 48 remaining bits cant be processed by the functional SMMs, there s no crossbars to send the datas in the functional SMMs caches.




Not sure, all they have to do is to use the 768MB for anything, even close to being useless, and technicaly it will be a 4GB card, the eventual weak point is the advertised 224GB/s bandwith wich can be proved to not being accurate, the 256bit bus claim cant be attacked since the bus is effectively 256 bit, it s just that 48 bits are almost useless for about anything else than said marketing.

I think you are absolutely on the right path. But I think there is crossbars to swap the data on the GPU. From what I experience and what I gather.

I think physically swapping the ram is inconvenient enough by itself and this is why they set the card up to use up the 3.5gb completely before hand.
See, I can get my GPU over 3300-3500 mb and its not a slide show. If the data had to go all the way back down the pice, nvidia would never allow that ram to be occupied. It's not reasonable and not what I expect is happenening

 

[Erenhardt]

ocre, my 7870 run Shadow of mordor quite nice with normal textures. On high it started a hitching every once couple of seconds. On ultra, it was stuttering. Its not like if you go 1mb beyond VRAM capacity you drop 75% of framerate, thou experience suffers a lot.

 

[n0x1ous]

PC Per just posted an explanation from Jonah Alben - GTX 970 only has 56 ROPS - there was a mistake in the reviewers guides and this hasn't turned up somehow.

http://www.pcper.com/reviews/Graphics-Cards/NVIDIA-Discloses-Full-Memory-Structure-and-Limitations-GTX-970

 

[5150Joker]

PC Per just posted an explanation from Jonah Alben - GTX 970 only has 56 ROPS - there was a mistake in the reviewers guides and this hasn't turned up somehow.

http://www.pcper.com/reviews/Graphics-Cards/NVIDIA-Discloses-Full-Memory-Structure-and-Limitations-GTX-970

As much as I like NVIDIA, I find the "oops our technical team messed up the guide" excuse b.s. The last 0.5 GB being 1/7th the speed would have given people pause about buying it and opting for a 980 or AMD card instead. I'm reading reports on Hard|OCP of people experiencing stutter once it hits >3.5 GB in Skyrim. Overall it's not a HUGE issue by any means but NVIDIA should have labeled it a 3.5 GB card and like PCPer suggested, call the other 0.5GB some kind of enhanced buffer or whatever.

 

[n0x1ous]

As much as I like NVIDIA, I find the "oops our technical team messed up the guide" excuse b.s.

As do I - they have had 4 months to correct this and only now fess up? Disappointing.

 

[Erenhardt]

But in the long term if the 7th port is fully busy, and is getting twice as many requests as the other port, then the other six must be only half busy, to match with the 2:1 ratio. So the overall bandwidth would be roughly half of peak.

So, Its 3,5GB at full speed or 4GB at half speed?

As much as I like NVIDIA, I find the "oops our technical team messed up the guide" excuse b.s.

Games rarely use that much vram. They probably bet no one will notice, and if someone will, it will be well too late in product lifecycle (EOL)

 

[Grooveriding]

Once third party entities start to test this, as they surely will now, we will learn even more. Right now all we have is nvidia trying to get out in front of it and spin it as best they can. Doesn't look good now. Doubtful they were so incompetent to confuse the number 64 with the number 56 and not realize it in the hundreds of reviews done.

I think end-users will ultimately lose on this. Even if this is enough to justify replacements or some monetary compensation in the minds of some buyers they'll never get it unless they are still in their return window.

 

[Granseth]

PC Per just posted an explanation from Jonah Alben - GTX 970 only has 56 ROPS - there was a mistake in the reviewers guides and this hasn't turned up somehow.

http://www.pcper.com/reviews/Graphics-Cards/NVIDIA-Discloses-Full-Memory-Structure-and-Limitations-GTX-970
(...)

Interesting, it seems like it they could just as well have dropped the last .5GB of VRAM.

The article also mentions the 660Ti with it's 2 GB of VRAM on 192 bit bus. It would have been interesting if someone would review the card again and see if it looses more performance now than it's bigger brothers now that Maxwell has driver priority. If so it should be a bad thing for the 970 in a 2 years time.

 

[jackstar7]

Does anyone think this will have actual market impact?

It seems like it will all be forgotten with the next release.

 

[destrekor]

PC Per just posted an explanation from Jonah Alben - GTX 970 only has 56 ROPS - there was a mistake in the reviewers guides and this hasn't turned up somehow.

http://www.pcper.com/reviews/Graphics-Cards/NVIDIA-Discloses-Full-Memory-Structure-and-Limitations-GTX-970

I think that goes to show that, while it was misleading, I don't think this issue is directly causing frame time differences.

Look at all the tests where they have forced full memory usage. These are states where you really wouldn't play the game at, period. The 980 can barely even handle those situations. The frame time difference between the 980 and 970 are unfair though, as the one article I saw, they only reported the worst 1% of measured times, and they were also looking at measurements when the average FPS was around 30fps, whereas the 980 was around, what 37-38? And that doesn't mention what the minimum and maximum were: the minimum could have been significantly higher, which helps the card keep frame time down and present more apparent smoothness.

At the end of the day, most of these cards can start developing stuttering and slow frame times when pushed to the brink and brought to play around 30fps. That isn't an ideal PC framerate. If it were CAPPED at 30 but the horsepower could push out 40 or more all the time, that would look fine, but barely managing to hit 30, it won't look good, ever.
And the 970 having a harder time than the 980 is much more the fault of having fewer SMMs than it is a memory constraint. It does appear to be apparent, a few percentage points, but it seems the drivers and OS are more smartly managing that memory for all conditions where the game would even be played at. Would it use 4GB of framebuffer if it could at that particular instance? Sure. But with smart management, if the game is at playable settings in the first place, it shouldn't be a noticeable issue, and really, a non-issue.

I'd like to see frame time analysis that shows the full time of measurement, as well as focuses on gameplay at playable settings. Nobody is going to max out settings and play at an average of 30fps. If they do, that's not a smart strategy to achieve smooth gameplay.

 

[destrekor]

So, Its 3,5GB at full speed or 4GB at half speed?

Did you read all of it? That is how it would be if they implemented it THAT way, but that is precisely what they averted and chose to implement it differently.

 

[NomanA]

Finally, it makes sense why GTX 970 has lower bandwidth in a portion of RAM, and also why 980M isn't affected by it.

This also settles the debate about Nai's CUDA tool utility in headless test conditions. The driver was making entire 4GB available to the tool which was then able to test the rates correctly when there was no contention on VRAM resources.

 

[destrekor]

Once third party entities start to test this, as they surely will now, we will learn even more. Right now all we have is nvidia trying to get out in front of it and spin it as best they can. Doesn't look good now. Doubtful they were so incompetent to confuse the number 64 with the number 56 and not realize it in the hundreds of reviews done.

I think end-users will ultimately lose on this. Even if this is enough to justify replacements or some monetary compensation in the minds of some buyers they'll never get it unless they are still in their return window.

The performance hasn't changed, so I think only those easily influenced are going to be returning. Honestly, I am still considering them to be my next cards. Why? The benchmark results are still accurate, and while I could be happy with the 290X, my situation would prefer the greater overclocking headroom with less heat in my case. That, and they are far more readily compatible with OS X Yosemite, so I can still have good performance in both OSes, it just might cost me a touch bit more. Oh well.

 

[wand3r3r]

Wow! Nvidia got caught red handed.

500 MB at a fraction of the speed isn't a nice surprise for SLI users and while every new game requires more VRAM.

the company did not fully disclose the missing L2 and ROP partition on the GTX 970, even if it was due to miscommunication internally.
...
Some applications are only aware of the first "pool" of memory and may only ever show up to 3.5GB in use for a game.

PCPer, aka NV's mouthpiece (undisclosed partnership with NV and FCAT after claiming it was theirs), downplays it somewhat.

Personally I think this is a joke and they should have just made it a 3.5 GB card without the deception. Clearly the remaining 500 MB can cause slowdowns with it's quite drastic speed reduction (obviously why they tried to aggressively avoid using it).

It's not often a company gets caught with such of a significant misleading product like this. I expect bumpgate lawsuits (not that I think they should/shouldn't).

So much for the claims that it's all fabricated.

 

[railven]

Interesting, it seems like it they could just as well have dropped the last .5GB of VRAM.

The article also mentions the 660Ti with it's 2 GB of VRAM on 192 bit bus. It would have been interesting if someone would review the card again and see if it looses more performance now than it's bigger brothers now that Maxwell has driver priority. If so it should be a bad thing for the 970 in a 2 years time.

I would totally go for that. I upgraded from sli gtx 660 ti because of stuttering in skyrim that was alleviated by disabling mods and dropping under 1.5gbs of vram usage.

 

[DooKey]

NV needs to take their lumps and reclass the 970 as a 3.5GB card and give all current buyers $50 checks or full refunds upon return of the card. Since the card seems to play games just fine I think most users would be glad to get a $50 rebate.

 

[Grooveriding]

This is interesting in that it also explains some of what mystified a reviewer at techreport and hardware.fr as to why the gtx 970 has such a sharp off drop-off in performance compared to the gtx 980 in certain situations.

Here's another reason the GTX 970 is slower than the GTX 980

For a while, I've thought I should drop you an email about some pixel fillrate numbers you use in the peak rates tables for GPUs. Actually, most people got those numbers wrong as Nvidia is not crystal clear about those kind of details unless you ask very specifically.
The pixel fillrate can be linked to the number of ROPs for some GPUs, but it’s been limited elsewhere for years for many Nvidia GPUs. Basically there are 3 levels that might have a say at what the peak fillrate is :

• The number of rasterizers
• The number of SMs
• The number of ROPs

On both Kepler and Maxwell each SM appears to use a 128-bit datapath to transfer pixels color data to the ROPs. Those appears to be converted from FP32 to the actual pixel format before being transferred to the ROPs. With classic INT8 rendering (32-bit per pixel) it means each SM has a throughput of 4 pixels/clock. With HDR FP16 (64-bit per pixel), each SM has a throughput of 2 pixels/clock.
On Kepler each rasterizer can output up to 8 pixels/clock. With Maxwell, the rate goes up to 16 pixels/clock (at least with the currently released Maxwell GPUs).


So the actual pixels/cycle peak rate when you look at all the limits (rasterizers/SMs/ROPs) would be :
GTX 750 : 16/16/16
GTX 750 Ti : 16/20/16
GTX 760 : 32/24/32 or 24/24/32 (as there are 2 die configuration options)
GTX 770 : 32/32/32
GTX 780 : 40/48/48 or 32/48/48 (as there are 2 die configuration options)
GTX 780 Ti : 40/60/48
GTX 970 : 64/52/64
GTX 980 : 64/64/64
Extra ROPs are still useful to get better efficiency with MSAA and so. But they don’t participate in the peak pixel fillrate.



That’s in part what explains the significant fillrate delta between the GTX 980 and the GTX 970 (as you measured it in 3DMark Vantage). There is another reason which seem to be that unevenly configured GPCs are less efficient with huge triangles splitting (as it’s usually the case with fillrate tests).

At the time they weren't fully clear on all the reasons it was happening. Now that the truth is out that there only 56 ROPS on the 970, not 64, that should explain it. It also explains the 4K weakness of the 970 vs 980 with how dependent on ROPS 4K performance generally is.

 

[SteveGrabowski]

What a load of crap. I wish I could return my 970 just on principle for Nvidia lying since September.

 

[SteveGrabowski]

NV needs to take their lumps and reclass the 970 as a 3.5GB card and give all current buyers $50 checks or full refunds upon return of the card. Since the card seems to play games just fine I think most users would be glad to get a $50 rebate.

I'd rather have a full refund and they can keep their $50 check. Boy did I screw up not getting a 290x.

 

[exar333]

This is interesting in that it also explains some of what mystified a reviewer at techreport and hardware.fr as to why the gtx 970 has such a sharp off drop-off in performance compared to the gtx 980 in certain situations.

Here's another reason the GTX 970 is slower than the GTX 980



At the time they weren't fully clear on all the reasons it was happening. Now that the truth is out that there only 56 ROPS on the 970, not 64, that should explain it. It also explains the 4K weakness of the 970 vs 980 with how dependent on ROPS 4K performance generally is.

Totally agree....

If I had know the 970 was a 3.5GB card, I don't think it would have changed my purchase decision last year, but it is still misinformation.

Knowing this doesn't somehow 'lower' my current framerates or anything, but its a pretty big deal that the card is 56 ROPS instead of 64, IMHO.

It seems unlikely that a large company like NV would have mistakenly used this miscalculation without someone bringing it up. I would totally understand if it was just a press-release at launch, but on a flagship product? Really?

I still think the 970 is a good deal vs. 980, but honestly there might be some people who would have rather purchased a 980 or 290x vs. 970 had they know the memory was actually less than the advertised 4GB (at full-speed) and ROP difference.

This story is getting more and more interesting.

 

[dacostafilipe]

This can be a problem for devs as their streaming engines should know about the second pool and adapt to it. Limiting total VRAM to 3.5Gb could even help performance.

I don't think this will stop me from getting a 970 in the next days, but it still bad for nVidia.

PS: Who's going the believe the story that they did not "see the error"? Really?

 

[wand3r3r]

NV needs to take their lumps and reclass the 970 as a 3.5GB card and give all current buyers $50 checks or full refunds upon return of the card. Since the card seems to play games just fine I think most users would be glad to get a $50 rebate.

Partially agreed.

I'm sure a large portion of the market won't care (they probably don't even know what 4k means i.e. not tech savvy), however there are certainly a lot of buyers thinking of SLI and higher resolutions with modern games assuming they bought a future-proof card.

The thread title could be updated to something about NV confirms 970 3.5 GB / 0.5 GB VRAM limitations and 56 ROPS etc. since it's no longer speculation.

 

[RampantAndroid]

What a load of crap. I wish I could return my 970 just on principle for Nvidia lying since September.

A class action suit should be possible for false advertising.

 

[sontin]

This is interesting in that it also explains some of what mystified a reviewer at techreport and hardware.fr as to why the gtx 970 has such a sharp off drop-off in performance compared to the gtx 980 in certain situations.

Here's another reason the GTX 970 is slower than the GTX 980

At the time they weren't fully clear on all the reasons it was happening. Now that the truth is out that there only 56 ROPS on the 970, not 64, that should explain it. It also explains the 4K weakness of the 970 vs 980 with how dependent on ROPS 4K performance generally is.

Just no.
The GTX970 can only process 52 pixel/clock because of the reduced SMM number. So even with 56 instead of 64 ROPs there doesnt exist a difference between the pixel throughput.

 

[SlowSpyder]

AMD and Nvidia advertise the R9 295 and Titan Z as 8GB and 12GB cards, when that really isn't the case in how the memory is used. My guess is nothing much will come of this, after all the price / performance is still the same regardless of what the specs were and are (with the possible exception of times when a game really could fill and use all 4GB of vram). But, I don't believe for a second that no one from Nvidia saw that sites were calling the GTX970 a 64 ROP GPU. It may have started out as an honest mistake, but I believe someone in the organization knew sooner or later but chose not to mention it.