Graphics Nvidia’s GTX970 has a rather serious memory allocation bug

As much difference as knowing that a 970 has 3.5Gb fast VRAM and 0.5GB of slower VRAM instead of the 4.0Gb VRAM I thought it had? The card still performs as well as it did when it was launched. It performs as well as it did last week, when you were perfectly happy with it. But people get so hung up on what they think they know...

Resale values always take a nose dive --such is the world of computer hardware. In two years' time we'll have 4k cards and the 970 and 980 will be obsolete as far as hard-core gamers are concerned. You don't buy gaming hardware for future proofing. You buy what allows you to play games now, at the cheapest price possible, because in two, three years you'll be upgrading again. The 970 is the wise gamer's choice.

Exactly. I am baffled how people aim to game at top 4k settings, on expensive 4k monitors, but then skimp on the GPU. If that is your ambition, buy the top-end card. Anything else is a compromise.

 

It only launched in September, so they've extended it a few weeks?

Yes, but that was already the case BEFORE this issue arose. I'm not knocking EVGA's step up programme, I'm knocking their attempt to wrap it up as an olive branch.

 

The discount they give is off the price you paid for 970. I'd be happy if they offer to take off RRP of a new 970.

Well, I never said I'm not happy with it now at any stage of this thread. I am looking to minimise problems in newer more VRAM demanding games.

I keep my hardware for more than 2 years. I kept my 580 for over 4 years, if it had 250MB of shuttering memory, I will not be able to play any of the games I played last year. The card aged well despite running with 1.5GB on 1440p. But evidences suggest 3.5-4GB is uncertain and 970 behavior is unpredictable above 3.5GB. I'll have to spend more money in 1 years time compared to my usual 3 to 4 years.

 

Doesn't that contradict your own statement? If a 250Mb card could keep up with games for 4 years, why would a 3.5Gb card suddenly start to struggle within a year? Unless games are getting exponentially more demanding as time goes on, in which case you'll fall behind the curve sooner or later.

 

My point is that a 3.5GB 970 should not struggle whereas I don't like uncertainty associated with 970's 3.5-4GB range. Keywords are bold for your reading pleasure

Hence my very original post about limiting VRAM to 3.5GB and forget about the 0.5GB. Skipping this whole problem directly and have a traditional memory controller, and that games are aware what actually is happening. (eg. comes around to games query VRAM size, which according to previous posts, might not be possible with a driver tweak)

 

Tempted myself as I need something with DP and can run WoW.

 

AMD takes aim at Nvidia's GTX 970 and cuts R9 290X price

http://hexus.net/tech/news/graphics/80130-amd-takes-aim-nvidias-gtx-970-cuts-r9-290x-price/

 

And then what when the game needs more than 3.5GB ?

 

my guess is the same as a 780 with only 3gb of ram, i think the 970's problem is when it uses the slower 512sb causing stuttering

 

This is my system specs with an Acer 4k panel.

In my unscientific testing, what happens is smooth framerate (@ 4k high or ultra settings) until the RAM goes much beyond 3700MB, then the GPU usage goes to 100%, and then depending on what is happening there are frame rate dips, or in some cases quite noticeable stutter (Star Citizen, SoM). When stationary however, smooth as butter. Does the same happen with 980s is the question, or does the full memory speed keep up with the load demand...

We agree the final 500MB runs at 1/8th speed, does that mean the whole of the memory then runs at 1/8th, or simply gets held up by it?

 

Sure the last 512MB runs at 1/8th the speed, but that's still 4x faster than system memory over the PCI-E bus.

Everything i have read indicates to me that segmenting the vRAM is far, far better than the alternatives, those being a card with 3GB on a 192bit bus, or accessing system memory that's orders of magnitude slower.

 

We've gone through this a few pages ago.

There potentially could be worse performance: (source)

If we treat it as a 3.5GB card from the beginning, everyone will be happy, it wouldn't have changed 970's performance at all.

 

Rather than going to system memory, I would assume the card attempts to rewrite to its own memory first, clearing any data not in active use.

Watching the usage in afterburner, it appears that the card tries to stay out of the second segment if it can. It goes right up to 3500MB almost immediately, for instance when I load the hanger module in SC, will not go above that, but then if I alt tab or start running around, it jolts then goes up to 3700MB or something. Then the stutters and dips occur.

 

But the potential for worse performance are massive if you look at the alternatives, like i said.

The alternatives are using system memory that's over 4x slower or dropping the card down to 3GB on a 192bit bus, not as yourself and others claim 3.5GB, AFAIK the technical limitations on what can or can't be disabled in the GPU means you can't just disable the MC that the 512MB of ram is attached to without losing other parts.

So no difference in how it handles 3.5/0.5 and 4GB then ?

 

I see more of you are coming into the stage of acceptance. We have had mocking and denial.

Anandtech say it better than I:

So no, not the same as having 4GB.

 

Sure you can, in the firmware write it's a 3.5GB card and that's that. The last 0.5GB is physically on the card but never used.

oh wait, that's not cost effective.

I also think I might have missed the part it says accessing the 0.5GB part will block access to 3.5GB completely. It's an XOR situation.

 

Conflating answer much ?

As you can see the answer i gave was in response to you saying "the card attempts to rewrite to its own memory first, clearing any data not in active use" To which i answered "no difference in how it handles 3.5/0.5 and 4GB then"

You can quote the Anandtech article until you're blue in the face, but you don't seem to even understand what he is saying, hear let me highlight the relevant parts for you...

In case you still don't understand what he's saying I'll try to put it in layman's terms.

IF, and it's a big IF the GPU tries to perform the same action on both segments, so if it attempted to write to both segments on the same clock cycle, or if it tried to read from both segments on the same cycle, then yes it would slow things down as it would have to wait 1 millisecond before it could read or write to the other segment.

Someone maybe able to correct me if I'm wrong, but while you maybe able to issue a firmware update that would report the card as a 3.5GB card you can't just turn it off, the only thing it would change is how the card reports it's self, not the way it works.

 

Its a big if that both sections may need writing to or reading from at the same time? What makes you say that?

 

I do wonder if some people would of actually paid out the extra £400 for 2 980s instead for for the 5-10% performance boost it offers have my douts on that one

 

Are you trying to suggesting that within the 4GB of data that the GPU is working on, that the GPU is using it's caching algorithms on, that it's going to read or write from both memory segments at the same time, within the same millisecond. That's even if it's a millisecond as i based that on the GPU clock, if the timings are based on the DRAM or crossbar frequency it would be even less.

People seem to be all in a fuss about this, but as i keep saying when you look at the alternatives this is by far the best solution (IMHO) That's unless you can come up with a better way of solving the following problem...
(Source)

So ask yourself when you look at this picture what would you disable, bearing in mind that you can't disable the MC without disabling the entire partition (the area in grey)