970's having performance issues using 4GB Vram - Nvidia investigating

[P0tenc]

Although I have not had any issues as yet, I am planning to buy a new high res monitor in the next few months.

My whole point of getting the GTX 970 was because of the 4GB frame buffer. Now I hear it's 3.5GB usable and 0.5GB snail speed.

Not happy

 

[Rangerjr1]

Hey man, you said it. IMO? if it was something they did not know about they would have made a statement to say "We really don't understand why this is happening and are looking into it".

But no. Instead they come along and tell you exactly why it's happening, because basically that's how the 970 uses its VRAM. They then post some benchmarks showing that it's not much worse than a 980....



What's the point of putting on, say, a 7990's box 6gb VRAM ! when we all know that it's really 3x3gb VRAM and you can only use 3gb of it because the other 3gb is just there to make the card run?

It's called marketing my friend. And once you understand that you'd be just as cynical and mistrusting as I am !

Yes, but the thing is we kind of know that from the get go, that dual GPU cards RAM dont add, there's nothing implying on the 970x that 500MB of your RAM is dog ****.

 

[FoxEye]

So much for 970 any day of the week in all 290 or 970 threads

I remember a thread from not so long ago, asking "Is the 970 the best VFM card ever made?"

99% of the responses were "It's the best ever!" People were so happy. We were young and naive. Now we know Jen Hsung is not the saviour he appeared to be. He did not come to spread amazing FPS to the masses. No, he just wanted to watch the world burn.

Damn him and his sub-optimal 500MB VRAM. Never forget.

 

[ALXAndy]

Yes, but the thing is we kind of know that from the get go, that dual GPU cards RAM dont add, there's nothing implying on the 970x that 500MB of your RAM is dog ****.

Had they launched it with 3.5gb VRAM no one would have bought it dude. It would then sound like it has far less going on than the 980. What Nvidia wanted people to think was that it was basically a 980 with a few things cut out that when overclocked matches the 980. It was the same with the 670. Hardly any one bought the 680, because IMO the 680 was only really there to make the 670 look like great value.

This is a lesson they learned from Fermi. They were down and out, almost out of business, then they come along with the 460 and *BANG* knock it out of the park. They sold like wildfire because they were cheap.

Since then they've left the kitchen sink well alone and have been making their profits selling lower end stuff for top money.

Pretty much genius, until you get caught lmao

 

[P0tenc]

Let me ask you this. Would you consider it acceptable to buy 8GB system Ram, just to find out later that 1GB of that Ram is running at an order of magnitude slower?

Someone needs to pay! Get the pitchforks!

 

[Dirk Diggler]

Can you do the same as I have done with my GTX 980

1. Use the same clock settings as I am running in GPUZ

2. Use the same bench settings in the pics below

3. Run the benchmark.

Spoiler

If your GTX 970 is working ok you should be looking for a score of 70.55 fps.

Score was 72.88, but as you can see it wouldn't use all the vRam.

 

[Midlife Crisis]

If this all had been common knowledge I probably would have still got a 970, just probably not a G1, one of the cheapest instead, or one of the cheaper
980's. I can't say I have had any problems with mine so far, but that itself is the issue you don't know when a game comes out that will be too much for it

 

[Mei]

Yes, but the thing is we kind of know that from the get go, that dual GPU cards RAM dont add, there's nothing implying on the 970x that 500MB of your RAM is dog ****.

yeh unless they go out of their way to prove it doesnt affect performance this wont go well for them lol
PR wise anyway
i doubt this is the only response we will see, this is just nvidia's safe response (the costs nothing response)

 

[james.miller]

Meh. If something comes out of this then fair play, but if not then meh. My 970 was a fantastic buy and given the choice id still go for the 970.

 

[Kaapstad]

Had they launched it with 3.5gb VRAM no one would have bought it dude. It would then sound like it has far less going on than the 980. What Nvidia wanted people to think was that it was basically a 980 with a few things cut out that when overclocked matches the 980. It was the same with the 670. Hardly any one bought the 680, because IMO the 680 was only really there to make the 670 look like great value.

This is a lesson they learned from Fermi. They were down and out, almost out of business, then they come along with the 460 and *BANG* knock it out of the park. They sold like wildfire because they were cheap.

Since then they've left the kitchen sink well alone and have been making their profits selling lower end stuff for top money.

Pretty much genius, until you get caught lmao

Imagine if NVidia had launched the 970 as it is equipped now but with 1gb of VRAM disabled in the bios so it was a 3gb card. It would not have taken long for someone to come up with a bios hack to turn the extra 1gb of VRAM on "just like the GTX 980". The best part about this is NVidia would have sold even more 970s as you can do this hack which makes them even better value for money and there would be no come back on NVidia.

For once NVidia may have missed a marketing trick !!!

 

[crinkleshoes]

It appears to only affect the 970 series...

Spoiler

Does that not also show that the throughput (bus / bit-width) has also been gimped on the 970?

On the first page of the thread... this:

Taken from Overclock.net
Quote:
Originally Posted by Serandur View Post
Quote:
Originally Posted by JohnLai View Post
=.= From the Nai's Benchmark, assuming if the allocation is caused by disabled of SMM units, and different bandwidth for each different gpus once Nai's Benchmark memory allocation reaches 2816MiBytes to 3500MiBytes range, I can only assume this is caused by the way SMM units being disabled.

Allow me to elaborate my assumption. As we know, there are four raster engines for GTX 970 and GTX 980.
Each raster engine has four SMM units. GTX 980 has full SMM units for each raster engine, so there are 16 SMM units.

GTX970 is made by disabling 3 of SMM units. What nvidia refused to told us is which one of the raster engine has its SMM unit being disabled.
I found most reviewers simply modified the high level architecture overview of GTX 980 diagram by removing one SMM unit for each three raster engine with one raster engine has four SMM unit intact.

First scenario
What if the first (or the second, third, fourth) raster engine has its 3 SMM units disabled instead of evenly spread across four raster engine?

Second scenario
Or, first raster engine has two SMM units disabled and second raster engine has one SMM unit disabled?

Oh, please do notice the memory controller diagram for each of the raster engine too. >.< If we follow the first scenario, definitely, the raster engine will not be able to make fully use of the memory controller bandwidth.
I agree that this is the most likely explanation, which would mean the issue is hardware-related and cannot be fixed. Here's an illustration of GM204:
** No hotlinking **
http://images.bit-tech.net/content_i...gtx980-17b.jpg

It does seem that each of the four 64-bit memory controllers corresponds with each of the four raster engines and in the same way that the 970's effective pixel fillrate has been demonstrated to be considerably lower than the 980's even though SMM cutting leaves the ROPs fully intact (http://techreport.com/blog/27143/her...an-the-gtx-980), the same situation may apply to bandwidth with Maxwell. However, the issue may be completely independent of which SMMs are cut and may simply relate to how many.

GM206's block diagram demonstrates the same raster engine to memory controller ratio/physical proximity:

http://cdn3.wccftech.com/wp-content/...ck-Diagram.jpg

I expect a cut-down GM206 part and even a GM200 part will exhibit the same issue as a result, it might be intrinsically tied to how Maxwell as an architecture operates. Cut down SMMs -> effectively mess up ROP and memory controller behavior as well as shaders and TMUs. I also don't think there's a chance in hell Nvidia were unaware of this, but I could be wrong.
Quote:
Originally Posted by JohnLai View Post
Yeah, it does seem the driver prioritizes on using 3.5Gb, otherwise the performance impact of using more than that would be too apparent.
But it would depend on how GTX 970 SMM being disabled.

64bit memory controller, total 4 memory controllers = 256 bit memory controller.


Assuming if there are 3 raster engines with each three has one SMM disabled leaving 1 raster engine with 4 SMM intact.
Mathematically ;
16 SMM = 256 bit = 4096 Mb
13 SMM = 208 bit = 3328 Mb

208 bit = effective width after disabling SMM with 256 bit being actual memory controller width.

I could be wrong on this oversimplify calculation. Still worry if one raster engine has three disabled SMM >.<

This behavior and throughput rates would agree with the above explanation, would it not?

That in order to limit the shader count, they have effectively just cut a portion of the chip, which includes the same portion of the memory bus.

So another point in that perhaps they have advertised a 256bit bus width but only delivered a 208bit bus width?

1664/2048 = 0.8125
208/256 = 0.8125

(From the 970 vs 980 runs above)
150/178 = 0.84



Not trying to add to the flames... just curious...

 

[P0tenc]

My 970 was a fantastic buy and given the choice id still go for the 970.

Probably a lot of people will feel the same as you do. I however would not have sold my 670's to upgrade to a 3.5GB card. The 4GB was the deciding factor for me. I would probably just have waited for AMDs new cards later this year.

 

[Kaapstad]

Score was 72.88, but as you can see it wouldn't use all the vRam.

I was not expecting GPUZ to show 4gb for the 970, this backs up what NVidia are saying that the 970 memory is split 3.5 and 0.5gb with GPUZ showing the 3.5gb part.

What were your minimums like as this will also highlight if the VRAM is holding you back.

72.88 for the average is a good score for a 970.

 

[Dirk Diggler]

** Edit *

Ignore this...

 

[queamin]

Meh. If something comes out of this then fair play, but if not then meh. My 970 was a fantastic buy and given the choice id still go for the 970.

If it cannot be fixed then the people who upgrade every couple of years who hold onto their cards longer could be seeing more problems down the road as games start to use more memory.

 

[ALXAndy]

16 SMM = 256 bit = 4096 Mb
13 SMM = 208 bit = 3328 Mb

*drums fingers....*

 

[varkanoid]

To be honest I dont care about the VRAM thing like some people get all het up about. I have NEVER EVER seen it be a problem when I play games.

The 970 is still a brilliant card for me.

 

[Dirk Diggler]

I was not expecting GPUZ to show 4gb for the 970, this backs up what NVidia are saying that the 970 memory is split 3.5 and 0.5gb with GPUZ showing the 3.5gb part.

What were your minimums like as this will also highlight if the VRAM is holding you back.

72.88 for the average is a good score for a 970.

Minimum was 46.16. Sorry I don't have a screen cap, I'm doing this in a rush!

 

[P0tenc]

If it cannot be fixed then the people who upgrade every couple of years who hold onto their cards longer could be seeing more problems down the road as games start to use more memory.

This is exactly the issue I am worried about. Plus monitors are getting higher pixel counts as we move away from 1080p gaming to 1440p.

 

[ALXAndy]

To be honest I dont care about the VRAM thing like some people get all het up about. I have NEVER EVER seen it be a problem when I play games.

The 970 is still a brilliant card for me.

The problem is it's already a problem. Shadow of Mordor is being limited to a set amount of VRAM.

Honestly, without sounding like a dick hole, what do you think is going to happen when GTAV releases?