GeForce GTX 970 Design Flaw Caps Video Memory Usage to 3.3 GB: Report

Blazzok said:
I just went and had a look at some reviews for both cards (980 and 970) where they tested them both at the same settings and with the same games, and tested the cards at 1080p as well as 4K. In all cases, the performance drop for both cards were mostly similar, so this might be why they did not pick it up. Perhaps they just never maxed out the VRAM on the 970? At 4K resolution though, I'm not sure how they didn't manage it..
Click to expand...

That's exactly my point though. It's a storm in a teacup if it doesn't have any significant effect.

So what if it has 1.75MB of cache instead of 2MB of cache. Is there anybody here who would not have bought the card because of that 0.25MB of cache? The same for 56ROPs vs 64ROPs. Those specifications have no real meaning unless they translate to a significant performance boost or impairment, and it seems that they don't in this case. So the specs are for the most part just for bragging rights...
 
Wyzak said:
That's exactly my point though. It's a storm in a teacup if it doesn't have any significant effect.
Click to expand...

Yep. I'm reading up a bit more now on review sites that did follow up benchmarking after this whole thing surfaced, and it seems that they are struggling to replicate the severe drops in frames with games. You mostly see the performance issue when you "artificially" fill the VRAM to capacity.

Here is one such article as well. Update3 has the most relevant information regarding the latest news on what the problem is.
For now it doesn't seem that returning your 970 would be the way to go, unless you bought them with the intent of playing 4K or something. If you want to play 4K though, Radeon cards seem to be the better option anyway.

http://www.guru3d.com/news-story/does-the-geforce-gtx-970-have-a-memory-allocation-bug.html

Oh well, I'm rambling. Looks like I won't be returning my 970. :)
 
No need to return the card/s, they are very good performers and they'll max most games easily.
The main issue here is the fact that Nvidia sold you an item with the wrong specs attached to it and then claimed they knew nothing about it and only after being forced into a corner they came forward and gave an explination.

I doubt anyone running 1080P or even 1440P will ever suffer any severe issues because of this, I can only see this happening at 4K with games that have incredible vram demand. For instance a game like The Witcher 3 might actually show some real performance drops at 4k because of this issue, although I doubt a 970 could run it smoothly at 4K an anyway.
 
The Joker said:
No need to return the card/s, they are very good performers and they'll max most games easily.
The main issue here is the fact that Nvidia sold you an item with the wrong specs attached to it and then claimed they knew nothing about it and only after being forced into a corner they came forward and gave an explination.

I doubt anyone running 1080P or even 1440P will ever suffer any severe issues because of this, I can only see this happening at 4K with games that have incredible vram demand. For instance a game like The Witcher 3 might actually show some real performance drops at 4k because of this issue, although I doubt a 970 could run it smoothly at 4K an anyway.
Click to expand...

^This. You summed it up quite well. Even though the cards will still perform awesomely, a lot of people will not be buying GTX970's now.
 
Blazzok said:
Here is one such article as well. Update3 has the most relevant information regarding the latest news on what the problem is.
For now it doesn't seem that returning your 970 would be the way to go, unless you bought them with the intent of playing 4K or something. If you want to play 4K though, Radeon cards seem to be the better option anyway.
Click to expand...

The 970 and 980 was always going to struggle at UHD due to it only have 4GB of RAM. So it would not have been a good idea to purchase that card if your intention was UHD.

Edit: 4K is a cinema resolution. UHD is the nearest equivalent on PC (4x FHD - 1920x1080).
 
Last edited:
Wyzak said:
The 970 and 980 was always going to struggle at UHD due to it only have 4GB of RAM. So it would not have been a good idea to purchase that card if your intention was UHD.

Edit: 4K is a cinema resolution. UHD is the nearest equivalent on PC (4x FHD - 1920x1080).
Click to expand...

True. I should rather have used 2160p I suppose, or the term UHD.
 
The Joker said:
Correct but true 4K monitors are now becoming available.
I'll be testing 1 next week.

http://www.amazon.co.uk/LG-31MU97-B-Flatron-Professional-WideScreen/dp/B00PQ9MYYA

Resolution: 4096x2160
Click to expand...

That will probably display with a black border on the edges due to it's inability to scale to the 16:9 ratio.

Blazzok: http://www.guru3d.com/news-story/does-the-geforce-gtx-970-have-a-memory-allocation-bug.html that's a nice article. It's just a shame that they refer to it as a big performance issue when the stats and even they themselves say otherwise:
As you can see the DRAM controller actually need to buddy up into the 7th L2 unit, that it the root cause of a big performance issue
Click to expand...

I'll try to find some time tonight to run that tester on my Gigabyte card and see if it also affects 3rd party cards (it probably will).
 
Wyzak said:
That will probably display with a black border on the edges due to it's inability to scale to the 16:9 ratio.

Blazzok: http://www.guru3d.com/news-story/does-the-geforce-gtx-970-have-a-memory-allocation-bug.html that's a nice article. It's just a shame that they refer to it as a big performance issue when the stats and even they themselves say otherwise:

I'll try to find some time tonight to run that tester on my Gigabyte card and see if it also affects 3rd party cards (it probably will).
Click to expand...

I'll post a review on it and see what it can do, unfortunately I only have a GTX980 and a GTX780Ti to test with, I have a funny feeling that the 780Ti will be faster @ 4K though.
 
Thing is, the quantifying fact is that nobody really has massive issues, dozens and dozens of media have tested the card with in-depth reviews like the ones here on my site. Replicating the stutters and stuff you see in some of the video's, well to date I have not been able to reproduce them unless you do crazy stuff, and I've been on this all weekend. Overall scores are good, and sure if you run out of memory at one point you will see perf drops. But then drop from 8 to like 4x AA right ?

To that end in the short amount of time we’ve had to work on this article we have also been working on cooking up potential corner cases for the GTX 970 and have so far come up empty, though we’re by no means done. Coming up with real (non-synthetic) gaming workloads that can utilize between 3.5GB and 4GB of VRAM while not running into a rendering performance wall is already a challenge, and all the more so when trying to find such workloads that actually demonstrate performance problems. This at first glance does seem to validate NVIDIA’s overall claims that performance is not significantly impacted by the memory segmentation, but we’re going to continue looking to see if that holds up. In the meantime NVIDIA seems very eager to find such corner cases as well, and if there are any they’d like to be able to identify what’s going on and tweak their heuristics to resolve them.

Ultimately we find ourselves going a full circle back to something NVIDIA initially said about the matter, which is that the performance impact of the GTX 970’s configuration is already baked into the results we have. After all, the configuration is not a bug or other form of unexpected behavior, and NVIDIA has been fully abstracting and handling the memory segments since the GTX 970’s initial launch. So while today’s revelation gives us a better understanding of how GTX 970 operates and what the benefits and drawbacks are, that information alone doesn’t change how the card behaves.
Click to expand...

Some quotes from that last article.
 
I did some tests on my card last night and there is definitely a slow-down as the amount of memory increases:
Nai's Benchmark, edited by VultureX
Device: GeForce GTX 970 (4.00 GB)
Memory Bus Width (bits): 256
Peak Theoretical DRAM Bandwidth (GB/s): 224.320000

Allocating Memory . . .
Chunk Size: 64 MiByte
Allocated 60 Chunks
Allocated 3840 MiByte
Benchmarking DRAM
DRAM-Bandwidth of Chunk no. 0 (0 MiByte to 64 MiByte):148.54 GByte/s
DRAM-Bandwidth of Chunk no. 1 (64 MiByte to 128 MiByte):147.84 GByte/s
DRAM-Bandwidth of Chunk no. 2 (128 MiByte to 192 MiByte):148.26 GByte/s
DRAM-Bandwidth of Chunk no. 3 (192 MiByte to 256 MiByte):147.93 GByte/s
DRAM-Bandwidth of Chunk no. 4 (256 MiByte to 320 MiByte):147.95 GByte/s
DRAM-Bandwidth of Chunk no. 5 (320 MiByte to 384 MiByte):147.45 GByte/s
DRAM-Bandwidth of Chunk no. 6 (384 MiByte to 448 MiByte):148.04 GByte/s
DRAM-Bandwidth of Chunk no. 7 (448 MiByte to 512 MiByte):148.75 GByte/s
DRAM-Bandwidth of Chunk no. 8 (512 MiByte to 576 MiByte):148.35 GByte/s
DRAM-Bandwidth of Chunk no. 9 (576 MiByte to 640 MiByte):148.34 GByte/s
DRAM-Bandwidth of Chunk no. 10 (640 MiByte to 704 MiByte):147.66 GByte/s
DRAM-Bandwidth of Chunk no. 11 (704 MiByte to 768 MiByte):147.99 GByte/s
DRAM-Bandwidth of Chunk no. 12 (768 MiByte to 832 MiByte):148.15 GByte/s
DRAM-Bandwidth of Chunk no. 13 (832 MiByte to 896 MiByte):150.08 GByte/s
DRAM-Bandwidth of Chunk no. 14 (896 MiByte to 960 MiByte):149.93 GByte/s
DRAM-Bandwidth of Chunk no. 15 (960 MiByte to 1024 MiByte):149.99 GByte/s
DRAM-Bandwidth of Chunk no. 16 (1024 MiByte to 1088 MiByte):149.65 GByte/s
DRAM-Bandwidth of Chunk no. 17 (1088 MiByte to 1152 MiByte):150.06 GByte/s
DRAM-Bandwidth of Chunk no. 18 (1152 MiByte to 1216 MiByte):149.97 GByte/s
DRAM-Bandwidth of Chunk no. 19 (1216 MiByte to 1280 MiByte):149.81 GByte/s
DRAM-Bandwidth of Chunk no. 20 (1280 MiByte to 1344 MiByte):149.31 GByte/s
DRAM-Bandwidth of Chunk no. 21 (1344 MiByte to 1408 MiByte):149.88 GByte/s
DRAM-Bandwidth of Chunk no. 22 (1408 MiByte to 1472 MiByte):149.89 GByte/s
DRAM-Bandwidth of Chunk no. 23 (1472 MiByte to 1536 MiByte):149.01 GByte/s
DRAM-Bandwidth of Chunk no. 24 (1536 MiByte to 1600 MiByte):150.76 GByte/s
DRAM-Bandwidth of Chunk no. 25 (1600 MiByte to 1664 MiByte):150.78 GByte/s
DRAM-Bandwidth of Chunk no. 26 (1664 MiByte to 1728 MiByte):150.25 GByte/s
DRAM-Bandwidth of Chunk no. 27 (1728 MiByte to 1792 MiByte):149.81 GByte/s
DRAM-Bandwidth of Chunk no. 28 (1792 MiByte to 1856 MiByte):151.42 GByte/s
DRAM-Bandwidth of Chunk no. 29 (1856 MiByte to 1920 MiByte):150.35 GByte/s
DRAM-Bandwidth of Chunk no. 30 (1920 MiByte to 1984 MiByte):151.04 GByte/s
DRAM-Bandwidth of Chunk no. 31 (1984 MiByte to 2048 MiByte):150.77 GByte/s
DRAM-Bandwidth of Chunk no. 32 (2048 MiByte to 2112 MiByte):150.99 GByte/s
DRAM-Bandwidth of Chunk no. 33 (2112 MiByte to 2176 MiByte):150.91 GByte/s
DRAM-Bandwidth of Chunk no. 34 (2176 MiByte to 2240 MiByte):149.97 GByte/s
DRAM-Bandwidth of Chunk no. 35 (2240 MiByte to 2304 MiByte):151.03 GByte/s
DRAM-Bandwidth of Chunk no. 36 (2304 MiByte to 2368 MiByte):150.76 GByte/s
DRAM-Bandwidth of Chunk no. 37 (2368 MiByte to 2432 MiByte):151.02 GByte/s
DRAM-Bandwidth of Chunk no. 38 (2432 MiByte to 2496 MiByte):150.43 GByte/s
DRAM-Bandwidth of Chunk no. 39 (2496 MiByte to 2560 MiByte):150.20 GByte/s
DRAM-Bandwidth of Chunk no. 40 (2560 MiByte to 2624 MiByte):151.21 GByte/s
DRAM-Bandwidth of Chunk no. 41 (2624 MiByte to 2688 MiByte):150.77 GByte/s
DRAM-Bandwidth of Chunk no. 42 (2688 MiByte to 2752 MiByte):150.39 GByte/s
DRAM-Bandwidth of Chunk no. 43 (2752 MiByte to 2816 MiByte):151.11 GByte/s
DRAM-Bandwidth of Chunk no. 44 (2816 MiByte to 2880 MiByte):151.20 GByte/s
DRAM-Bandwidth of Chunk no. 45 (2880 MiByte to 2944 MiByte):27.30 GByte/s
DRAM-Bandwidth of Chunk no. 46 (2944 MiByte to 3008 MiByte):27.29 GByte/s
DRAM-Bandwidth of Chunk no. 47 (3008 MiByte to 3072 MiByte):27.30 GByte/s
DRAM-Bandwidth of Chunk no. 48 (3072 MiByte to 3136 MiByte):27.30 GByte/s
DRAM-Bandwidth of Chunk no. 49 (3136 MiByte to 3200 MiByte):27.30 GByte/s
DRAM-Bandwidth of Chunk no. 50 (3200 MiByte to 3264 MiByte):27.31 GByte/s
DRAM-Bandwidth of Chunk no. 51 (3264 MiByte to 3328 MiByte):27.30 GByte/s
DRAM-Bandwidth of Chunk no. 52 (3328 MiByte to 3392 MiByte):27.29 GByte/s
DRAM-Bandwidth of Chunk no. 53 (3392 MiByte to 3456 MiByte):27.30 GByte/s
DRAM-Bandwidth of Chunk no. 54 (3456 MiByte to 3520 MiByte):27.30 GByte/s
DRAM-Bandwidth of Chunk no. 55 (3520 MiByte to 3584 MiByte): 7.79 GByte/s
DRAM-Bandwidth of Chunk no. 56 (3584 MiByte to 3648 MiByte): 7.79 GByte/s
DRAM-Bandwidth of Chunk no. 57 (3648 MiByte to 3712 MiByte): 7.78 GByte/s
DRAM-Bandwidth of Chunk no. 58 (3712 MiByte to 3776 MiByte): 7.78 GByte/s
DRAM-Bandwidth of Chunk no. 59 (3776 MiByte to 3840 MiByte): 7.79 GByte/s
Benchmarking L2-Cache
L2-Cache-Bandwidth of Chunk no. 0 (0 MiByte to 64 MiByte):431.78 GByte/s
L2-Cache-Bandwidth of Chunk no. 1 (64 MiByte to 128 MiByte):431.44 GByte/s
L2-Cache-Bandwidth of Chunk no. 2 (128 MiByte to 192 MiByte):431.54 GByte/s
L2-Cache-Bandwidth of Chunk no. 3 (192 MiByte to 256 MiByte):431.59 GByte/s
L2-Cache-Bandwidth of Chunk no. 4 (256 MiByte to 320 MiByte):431.74 GByte/s
L2-Cache-Bandwidth of Chunk no. 5 (320 MiByte to 384 MiByte):431.54 GByte/s
L2-Cache-Bandwidth of Chunk no. 6 (384 MiByte to 448 MiByte):431.74 GByte/s
L2-Cache-Bandwidth of Chunk no. 7 (448 MiByte to 512 MiByte):431.74 GByte/s
L2-Cache-Bandwidth of Chunk no. 8 (512 MiByte to 576 MiByte):431.73 GByte/s
L2-Cache-Bandwidth of Chunk no. 9 (576 MiByte to 640 MiByte):431.61 GByte/s
L2-Cache-Bandwidth of Chunk no. 10 (640 MiByte to 704 MiByte):431.63 GByte/s
L2-Cache-Bandwidth of Chunk no. 11 (704 MiByte to 768 MiByte):431.67 GByte/s
L2-Cache-Bandwidth of Chunk no. 12 (768 MiByte to 832 MiByte):431.50 GByte/s
L2-Cache-Bandwidth of Chunk no. 13 (832 MiByte to 896 MiByte):431.97 GByte/s
L2-Cache-Bandwidth of Chunk no. 14 (896 MiByte to 960 MiByte):431.82 GByte/s
L2-Cache-Bandwidth of Chunk no. 15 (960 MiByte to 1024 MiByte):431.48 GByte/s
L2-Cache-Bandwidth of Chunk no. 16 (1024 MiByte to 1088 MiByte):431.61 GByte/s
L2-Cache-Bandwidth of Chunk no. 17 (1088 MiByte to 1152 MiByte):431.74 GByte/s
L2-Cache-Bandwidth of Chunk no. 18 (1152 MiByte to 1216 MiByte):431.61 GByte/s
L2-Cache-Bandwidth of Chunk no. 19 (1216 MiByte to 1280 MiByte):431.54 GByte/s
L2-Cache-Bandwidth of Chunk no. 20 (1280 MiByte to 1344 MiByte):431.56 GByte/s
L2-Cache-Bandwidth of Chunk no. 21 (1344 MiByte to 1408 MiByte):431.63 GByte/s
L2-Cache-Bandwidth of Chunk no. 22 (1408 MiByte to 1472 MiByte):431.86 GByte/s
L2-Cache-Bandwidth of Chunk no. 23 (1472 MiByte to 1536 MiByte):431.71 GByte/s
L2-Cache-Bandwidth of Chunk no. 24 (1536 MiByte to 1600 MiByte):431.57 GByte/s
L2-Cache-Bandwidth of Chunk no. 25 (1600 MiByte to 1664 MiByte):431.71 GByte/s
L2-Cache-Bandwidth of Chunk no. 26 (1664 MiByte to 1728 MiByte):431.61 GByte/s
L2-Cache-Bandwidth of Chunk no. 27 (1728 MiByte to 1792 MiByte):431.69 GByte/s
L2-Cache-Bandwidth of Chunk no. 28 (1792 MiByte to 1856 MiByte):431.65 GByte/s
L2-Cache-Bandwidth of Chunk no. 29 (1856 MiByte to 1920 MiByte):431.52 GByte/s
L2-Cache-Bandwidth of Chunk no. 30 (1920 MiByte to 1984 MiByte):431.84 GByte/s
L2-Cache-Bandwidth of Chunk no. 31 (1984 MiByte to 2048 MiByte):431.35 GByte/s
L2-Cache-Bandwidth of Chunk no. 32 (2048 MiByte to 2112 MiByte):431.50 GByte/s
L2-Cache-Bandwidth of Chunk no. 33 (2112 MiByte to 2176 MiByte):431.54 GByte/s
L2-Cache-Bandwidth of Chunk no. 34 (2176 MiByte to 2240 MiByte):431.42 GByte/s
L2-Cache-Bandwidth of Chunk no. 35 (2240 MiByte to 2304 MiByte):431.69 GByte/s
L2-Cache-Bandwidth of Chunk no. 36 (2304 MiByte to 2368 MiByte):431.78 GByte/s
L2-Cache-Bandwidth of Chunk no. 37 (2368 MiByte to 2432 MiByte):431.71 GByte/s
L2-Cache-Bandwidth of Chunk no. 38 (2432 MiByte to 2496 MiByte):431.56 GByte/s
L2-Cache-Bandwidth of Chunk no. 39 (2496 MiByte to 2560 MiByte):431.73 GByte/s
L2-Cache-Bandwidth of Chunk no. 40 (2560 MiByte to 2624 MiByte):431.73 GByte/s
L2-Cache-Bandwidth of Chunk no. 41 (2624 MiByte to 2688 MiByte):431.59 GByte/s
L2-Cache-Bandwidth of Chunk no. 42 (2688 MiByte to 2752 MiByte):431.69 GByte/s
L2-Cache-Bandwidth of Chunk no. 43 (2752 MiByte to 2816 MiByte):431.74 GByte/s
L2-Cache-Bandwidth of Chunk no. 44 (2816 MiByte to 2880 MiByte):431.52 GByte/s
L2-Cache-Bandwidth of Chunk no. 45 (2880 MiByte to 2944 MiByte):95.21 GByte/s
L2-Cache-Bandwidth of Chunk no. 46 (2944 MiByte to 3008 MiByte):95.21 GByte/s
L2-Cache-Bandwidth of Chunk no. 47 (3008 MiByte to 3072 MiByte):95.20 GByte/s
L2-Cache-Bandwidth of Chunk no. 48 (3072 MiByte to 3136 MiByte):95.22 GByte/s
L2-Cache-Bandwidth of Chunk no. 49 (3136 MiByte to 3200 MiByte):95.21 GByte/s
L2-Cache-Bandwidth of Chunk no. 50 (3200 MiByte to 3264 MiByte):95.21 GByte/s
L2-Cache-Bandwidth of Chunk no. 51 (3264 MiByte to 3328 MiByte):95.20 GByte/s
L2-Cache-Bandwidth of Chunk no. 52 (3328 MiByte to 3392 MiByte):95.22 GByte/s
L2-Cache-Bandwidth of Chunk no. 53 (3392 MiByte to 3456 MiByte):95.21 GByte/s
L2-Cache-Bandwidth of Chunk no. 54 (3456 MiByte to 3520 MiByte):95.22 GByte/s
L2-Cache-Bandwidth of Chunk no. 55 (3520 MiByte to 3584 MiByte): 7.32 GByte/s
L2-Cache-Bandwidth of Chunk no. 56 (3584 MiByte to 3648 MiByte): 7.32 GByte/s
L2-Cache-Bandwidth of Chunk no. 57 (3648 MiByte to 3712 MiByte): 7.32 GByte/s
L2-Cache-Bandwidth of Chunk no. 58 (3712 MiByte to 3776 MiByte): 7.33 GByte/s
L2-Cache-Bandwidth of Chunk no. 59 (3776 MiByte to 3840 MiByte): 7.32 GByte/s
Press any key to continue . . .
Click to expand...
 
Wyzak said:
I did some tests on my card last night and there is definitely a slow-down as the amount of memory increases:
Click to expand...

Very interesting to see that the slow down starts at the 3GB mark already. It must be to do with the memory controllers sharing the L2 cache unit (see the Anandtech article that I linked earlier in the thread).
 
StaggerLee said:
Very interesting to see that the slow down starts at the 3GB mark already. It must be to do with the memory controllers sharing the L2 cache unit (see the Anandtech article that I linked earlier in the thread).
Click to expand...

I'm not sure how this application really works and whether it's truly a valid test to run it in Windows. If you think about it obviously some of the video memory will already be in use when this application starts consuming memory like it's going out of fashion. That may be why it starts sooner than expected. I noticed some erratic behaviour with the application as well.
 
Wyzak said:
I'm not sure how this application really works and whether it's truly a valid test to run it in Windows. If you think about it obviously some of the video memory will already be in use when this application starts consuming memory like it's going out of fashion. That may be why it starts sooner than expected. I noticed some erratic behaviour with the application as well.
Click to expand...

This was also mentioned in the Guru3D article about this and that is a valid concern. I think they recommended turning off windows aero and then turning off your screen to minimise the load on the card while you do the test. Maybe that would make a difference and would be interesting to compare results generated in that way.
 
StaggerLee said:
This was also mentioned in the Guru3D article about this and that is a valid concern. I think they recommended turning off windows aero and then turning off your screen to minimise the load on the card while you do the test. Maybe that would make a difference and would be interesting to compare results generated in that way.
Click to expand...

Yeah I didn't do that :D
 
That vid was funny as hell now i'm glad i never decided to go for that 970 galax. Joker if your reading any of this does the Galax 970 have coil whine and does it also have this problem?
 
AlienTaker said:
That vid was funny as hell now i'm glad i never decided to go for that 970 galax. Joker if your reading any of this does the Galax 970 have coil whine and does it also have this problem?
Click to expand...
I wonder if the story he's telling is as funny? [emoji14]
 
You must log in or register to reply here.
Back
Top