Nvidia explains the GTX 970's memory 'problems'

[azrael-]

<SNIP>
Self-flagellation aside, it's strange that it's a 3.5GB : 0.5GB split - the last time I saw that was on the Dell D620 I bought from eBay. Despite having 4GB installed, I could only ever use 3.5GB, wonder if it's the same kind of addressing deal here?
<SNIP>

Not to dilute this thread with unrelated information, but no it's not. Those 512MB you've "lost" is address space to which e.g. the PCI/AGP/PCIe et al address space is mapped. Otherwise the system wouldn't know how to address the memory of components that aren't using main memory. You won't necessarily lose 512MB either. Can be more, can be quite a bit less. For instance my old Windows XP installation (which is very rarely used today, if ever) maps my GTX 670's memory into the upper 256 MB. That's about how much I lose of the 4 GB accessible to 32 bit Windows XP.

[shaithis]

However a blatant attempt to mislead customers matters a lot in my book. They supplied incorrect specifications to review sites, lets give them the benefit of the doubt and say this was a mistake

Call me naive but when I am buying a graphics card performance is my number one metric, followed by features and price.
I have never bought a gfx card based on how many SM units it has.....I go by performance differential over my current card.

Now, I am far from your average user. Did nVidia gain anything from this? Did it skew benchmarks? Did it make the card seem faster than it would be when you got it home? Did anyone even notice?

Storm in a teacup.

[Jowsey]

Call me naive but when I am buying a graphics card performance is my number one metric, followed by features and price.
I have never bought a gfx card based on how many SM units it has.....I go by performance differential over my current card.

Now, I am far from your average user. Did nVidia gain anything from this? Did it skew benchmarks? Did it make the card seem faster than it would be when you got it home? Did anyone even notice?

Storm in a teacup.

You can't argue that the performance isn't fantastic.

However, did a they lie to us? Yes
Was it not just a buffed up half lie? No, they lied on official spec sheets
Do I trust them as much as I used to (...as far as you 'trust' a company trying to sell something to you)?No

My current card is a GTX760 which performs great at 1080p and I don't have intentions of purchasing a higher resolution monitor so I will probably just stick with the 760 till it dies. But when I do decide to buy a new card, I will think a little more carefully and not just base it purely benchmark performances

[CAT-THE-FIFTH]

There is a long thread over on OcUK about this:
http://forums.overclockers.co.uk/sho...651061&page=70

If you look on that page someone did some testing on their card and if it exceeds 3.5GB usage it becomes very stuttery. He is not the only person to see something similar.

It appears the driver is doing what it can to not exceed 3.5GB. This all smacks of marketing wanting to have the GTX970 as a 4GB card.

[watercooled]

As has been said, 'no impact on performance' isn't necessarily true as the main problem caused by this is apparently stuttering, which is generally not picked up in FPS measurements. Reviewers by and large don't seem to have bothered with frame time measurements for the Maxwell cards, which presumably would have picked it up as the quick tests by Extremetech which I linked begin to show. And of course there are the subjective complaints of stuttering which probably led to the investigation in the first place - this seems to have been spotted first by end users noticing performance issues, not by reviewers.

Anyway, some more on the details of the problem:
Something which hasn't been made clear AFAICT, is how accessing this portion of VRAM impacts the adjacent memory channel. They list 196GB/sec for the big pool, 28GB/sec for the slow pool, but as Anandtech explains, not simultaneously. I guess they're prioritising the fast pool but unless the heuristics are spot on, this means at either extreme accessing the slow pool also drags down the bandwidth of the adjacent port, or the bandwidth of the slow pool is crippled, perhaps even below PCIe speed as it's not far above to start with. I.e rather than just the upper 512m being impacted, it seems this could have implications for anything in the top GB (and potentially impact far more than that depending on how data is striped).

[shaithis]

Only really effecting performance when people are finding ways to force usage over 3.5GB. It seems the drivers do a good job of stopping usage there, hence reviewers didn't pick it up I guess.

Does seem a waste of 0.5GB of DDR5 per card. i wonder how much they could have knocked off the price if they only came with 3.5GB?

As for marketing reasons.....I ain't so sure....I would have thought that a 3.5GB 970 and a 4GB 980 would help separate the models more and potentially allow them to sell the 980 for a tad more.

[ekamdeep]

happy with my r9 290 dd
was thinking about gtx 970 at first
now i think i made the right decision

[Corky34]

Does seem a waste of 0.5GB of DDR5 per card. i wonder how much they could have knocked off the price if they only came with 3.5GB?

I think i read it's because disabling one memory controller means disabling the whole ROP/L2, so you would end up with a 3Gb card with less of other stuff that i can't remember.

Maybe in the future we will have even finer grain control of what is enabled or disabled in a GPU.

[HalloweenJack]

it would seem that those with GTX 970 with Hynix ram are suffering (or more vocal) than those with Samsung ram

*cough* R9 290 Elpida *cough*

[watercooled]

I'm not sure why RAM manufacturer would have an impact here. Maybe there are simply more cards produced with Hynix RAM?

[kalniel]

it would seem that those with GTX 970 with Hynix ram are suffering (or more vocal) than those with Samsung ram

*cough* R9 290 Elpida *cough*

Make of ram is irrelevant, it's the memory controller design that's causing this. But I think this is a non-issue. There's going to be so few situations where 3.5GB is not enough but 4GB is fine.

[HalloweenJack]

Make of ram is irrelevant, it's the memory controller design that's causing this. But I think this is a non-issue. There's going to be so few situations where 3.5GB is not enough but 4GB is fine.

no its no irrelevant - or do you forget the R9 290 with elpida ram *OMG MY CARD IS DEAD* problem??

Hynix ram users on the 970`s make up for about 80% of forum users who are vocal about stuttering in games when the card is being hammered

and please tell those its a `non issue` when games are stuttering on a 970 but on a 980 they are fine.

please , im sure they will bathe in your words of wisdom

[kalniel]

and please tell those its a `non issue` when games are stuttering on a 970 but on a 980 they are fine.

Because the 980 memory crossbar has full speed access to the full 4GB, as explained in the article.

The worrying (albeit not surprising to anyone who knows nvidia) bit is this:

we have learned that Nvidia has known all along that the information passed along to reviewers - ROP counts, L2 cache, etc. - has been wrong... and has done nothing about it until forced to do so when speculation grew too rife.

[spl]

Problem is games like Shadow of Mordor and Advanced Warfare that will use as much video RAM is available to them. I wonder if it would be beneficial then to appear as a 3.5GB card to those games? With dual-channel DDR3-1600 (or triple DDR3-1066) you've got 25.6GB/s total bandwidth on the system memory, so it could actually be faster for textures etc to spill over into system memory than use that last 512MB GDDR5 running at 20GB/s. I think either is going to hurt performance a lot more than nVIDIA suggests though.

[crossy]

no its no irrelevant - or do you forget the R9 290 with elpida ram *OMG MY CARD IS DEAD* problem?? Hynix ram users on the 970`s make up for about 80% of forum users who are vocal about stuttering in games when the card is being hammered

Feel free to shoot me down in flames, but I think the 290/Elpida and the 970/Hynix cases aren't comparable. A dead card (Elpida) is an unarguable case for an RMA, on the other hand a card with a visual stutter is going to be a hard sell for RMA unless you've got a pretty understanding merchant. Certainly if I was Scan/eBuyer/etc then I'd be very wary of doing refunds, since I'm sure that there's a proportion of folks who'd cry wolf over issues to be able to get started on the road to a 980, now that the 970 is less of the bargain that it looked to be a week ago.

Can someone indulge my curiosity - how common is it for a current game title to request that full 4GB? Or are games these days programmed to ram (no pun intended) as much texture info into the GDDR as possible in a caching arrangement. I'm not a 970 owner (and after this I've no intention of being) so it is really just idle curiosity on my part.

[Corky34]

With dual-channel DDR3-1600 (or triple DDR3-1066) you've got 25.6GB/s total bandwidth on the system memory, so it could actually be faster for textures etc to spill over into system memory than use that last 512MB GDDR5 running at 20GB/s. I think either is going to hurt performance a lot more than nVIDIA suggests though.

Isn't the maximum throughput for a x16 PCI-e slot around 15 GB/s ?