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

100% correct.

Then again they wouldn't have sold as many, and I'm sure more people would have ended up just sticking to their older cards or upgraded to the 980.

Agh I think they would have, AMD just doesn't have anything to compete atm. So regardless of whether it was 3.5gb or 4gb they would have sold. But if anything they would have saved face by making it 3.5gb.
 
100% correct.

Then again they wouldn't have sold as many, and I'm sure more people would have ended up just sticking to their older cards or upgraded to the 980.

I'm not so sure that they would have sold any less. Any hardware enthusiast worth his salt would have looked at more specs than just the memory and would probably have bought the card based on reliable benchmarks rather than specs alone. I think the card would have done just as well. It's silly to speculate I suppose, we can't ever know what might have been, but I still think the issue is a storm in teacup.
 
I'm not so sure that they would have sold any less. Any hardware enthusiast worth his salt would have looked at more specs than just the memory and would probably have bought the card based on reliable benchmarks rather than specs alone. I think the card would have done just as well. It's silly to speculate I suppose, we can't ever know what might have been, but I still think the issue is a storm in teacup.

Blamless from OC.net says it best:

Because the GTX 970 needed 4GiB of RAM on paper to look appealing, but they couldn't allow it it to be as fast as the 980 or it would cannibalize too much of the more expensive card's market share. The partial disable "feature" is a feature for NVIDIA to enable them to produce more granular segmentation.

They could have disabled more SMs and just left the ROM/MC cluster fully enabled, but that would have resulted in the 760 ti's getting too close in shader performance and made the 970 slower than it is now, because it wasn't ROP limited in the first place.

The only other option would have been to disable the ROP/MC cluster entirely, which would have made the 970 a 3GiB card.

What NVIDIA did to the 970 was the best way they could satisfy the segmentation they want to enforce. Any other option would have resulted in either a slower card at the same price that would not have sold as well, or just a lower clocked 980 that would have sold too well, to the detriment of the more profitable part.

None of this excuses the distribution of false information, mistaken or otherwise, but it does explain why the GTX 970 is configured the way it is.
 
Ye look, I think they proper screwed themselves. It will be even worse if that court case goes bad. Interesting times ahead I reckon.
 
Good Grief! how does 970 SLI not handle this, is it purely cause of the memory issue?

Part of the problem is that SLI requires you to have mirrored memory (Crossfire, through Mantle, DX12 or XDMA doesn't have to work this way anymore). When you're syncing memory over the SLI bus, that 512MB second pool is written to from the system at super-slow speeds and then it gets synced over super-slowly. The effect is really exacerbated in SLI and it'll take a lot of driver improvements to make it less impactful on system performance.

Then again they wouldn't have sold as many, and I'm sure more people would have ended up just sticking to their older cards or upgraded to the 980.

PC Perspective's Sebastian Peak made the same point that I've been mentioning in other forums online for a while. Had Nvidia called this a 3.5GB card with a 512MB cache, people would have accepted the realities of putting all of the memory on the card at a lower price point and would have said, "Well I don't game at 4K so I don't need the full 4GB. Look at how much performance I'm getting from this card at $200 less than a 980!"

Nvidia missed a chance to use Maxwell's memory arrangement to its advantage in the card's marketing and it would have been even more successful than before. "Does your R9 290X have a special 512MB cache for infrequently used game data? No? Suck it AMD!"
 
Part of the problem is that SLI requires you to have mirrored memory (Crossfire, through Mantle, DX12 or XDMA doesn't have to work this way anymore). When you're syncing memory over the SLI bus, that 512MB second pool is written to from the system at super-slow speeds and then it gets synced over super-slowly. The effect is really exacerbated in SLI and it'll take a lot of driver improvements to make it less impactful on system performance.



PC Perspective's Sebastian Peak made the same point that I've been mentioning in other forums online for a while. Had Nvidia called this a 3.5GB card with a 512MB cache, people would have accepted the realities of putting all of the memory on the card at a lower price point and would have said, "Well I don't game at 4K so I don't need the full 4GB. Look at how much performance I'm getting from this card at $200 less than a 980!"

Nvidia missed a chance to use Maxwell's memory arrangement to its advantage in the card's marketing and it would have been even more successful than before. "Does your R9 290X have a special 512MB cache for infrequently used game data? No? Suck it AMD!"

I think if they went ur cached route, this would have caused even more issues. Think about it, its referred to a cached memory but if u dare access that memory ur performance goes back to the dark ages. This would have been marketing suicide.
 
Blamless from OC.net says it best:

Well, yes. Pretty much.

Although, what he doesn't mention here:
Any other option would have resulted in either a slower card at the same price that would not have sold as well, or just a lower clocked 980 that would have sold too well, to the detriment of the more profitable part.
is that if they had taken this approach, the 970 would have been a more expensive card in the first place, making it pointless to put on the market anyway. That is the beauty of the higher granularity, the highest possible performance for the best price.

The mistakes that were made were marketing mistakes. They distributed incorrect info and they sold the card as a 4Gig card and not a 3.5Gig card with an extra backup wad of memory (as Wesley mentioned).
 
Well, yes. Pretty much.

Although, what he doesn't mention here:

is that if they had taken this approach, the 970 would have been a more expensive card in the first place, making it pointless to put on the market anyway. That is the beauty of the higher granularity, the highest possible performance for the best price.

The mistakes that were made were marketing mistakes. They distributed incorrect info and they sold the card as a 4Gig card and not a 3.5Gig card with an extra backup wad of memory (as Wesley mentioned).

Im sure AMD will milk this for some time to come. I wonder if reference of this will be seen in the R300 series of cards.
 
I think if they went ur cached route, this would have caused even more issues. Think about it, its referred to a cached memory but if u dare access that memory ur performance goes back to the dark ages. This would have been marketing suicide.

That's not a possibility, actually. Nvidia has sold just over 13 cards with mismatched memory pools in the past six years and their algorithms for deciding what data goes into the second pool and what does not is quite mature in Maxwell's case.

The issues that persist today are with a very limited set of games that hammer the VRAM and fill it up as much as possible. Titles like Hitman: Absolution, Crysis 3, or Shadow of Mordor haven't been included in Nvidia's algorithms for Maxwell because they haven't had to deal with this yet. The process of optimisation is ongoing and will be completed much sooner because they can take a lot of what they learned from Kepler and apply it here.

Even then, there will still be games that don't behave properly with that second pool. I expect them to not do any optimisation for Skyrim or Max Payne 3. Or GTA V for that matter, because the issue with those titles is in the game engine itself, not the drivers. In fact, you can take the GTX 660Ti with 2GB of VRAM and compare it to the same card with 3GB of VRAM in tests with Max Payne 3 - the 2GB card will only ever try to load up to 1.5GB and will stutter if you use more than that, while the 3GB card has no issues of that nature.

A lot of people don't understand that Maxwell's capability to turn off L2 cache and selectively pace memory pools to different controllers is new territory even to Nvidia. The GTX 970 is effectively a 192-bit card most of the time, but it can occasionally be a 256-bit card as well. All the associated strange behaviour as a result of the Jekyll/Hyde setup here will be new things they've never had to deal with before.

Also note that this capability foreshadows something they're almost certainly working on to improve power efficiency. Memory dies need to be fed with power from the memory bus, there's no way to selectively turn them off to save power. With Maxwell, it could be possible that they've designed a way to selectively turn off memory and L2 cache associated with the card and it may be something they can only do in DX12, since it will scale performance up and down with the number of compute resources becoming available (one of the benefits of GPGPU and OpenCL). They'll have something very much like this for GPUs based on Pascal in the future.
 
Last edited:
Wow very interesting read! Thanks for taking the time to write that up!

So as long as NVidia can keep tweaking their drivers, new games should be able to manage this issue? Assuming there isn't a problem with the engine itself.
 
I found this to be some what interesting and think I'm still gonna go for this card.




 
Last edited:
^^ without watching that video, it is WELL worth it. Can play everything maxxed out!
I'm loving mine, NO regrets at all! Only regret is not purchasing it earlier!
 
^^ without watching that video, it is WELL worth it. Can play everything maxxed out!
I'm loving mine, NO regrets at all! Only regret is not purchasing it earlier!

Figured as much, but surely since this whole debacle it has stirred up some price drops since it's release (and media controversy)?


Edit: Funny enough the benchmarks show it performs better at 1080p as opposed to 1440p.
 
Last edited:
Back
Top