970 memory allocation issue revisited

We can't say anything at all definitively, other than this is a really difficult issue to "prove" since plenty of people are not seeing any performance issues under normal game play circumstances.

I can point you back to my post from this morning on the previous page. I did some testing with Titanfall last night at 1440 with everything maxed out and 2xMSAA. The game topped out at 3540MB of VRAM usage and wouldn't go higher than that, but it ran smooth as silk through a full match. Increasing the resolution and AA beyond that didn't make the game look any better and only cause frame dips and increased frame times, but the game never "hitched" or "stuttered" like it was having a memory bottleneck. What I saw was a gradual degradation of performance as I pushed things up, which is exactly what I would expect to see. Since it didn't make the game look any better, who cares?

 

I thought that the amount of RAM needed on a graphic card is directly related to the size of the textures used at different resolutions. So the higher the resolution, the more space is needed to keep the textures. Any comment?

Since I prefer higher refresh rate (120hz) over higher resolution (which would be anything over 1440p), will I ever see any problems? Namely, can there theoretically be any games in the future which will require over 3,3 GB of RAM on GPUs for resolutions up to 1080p or 1440p?

If I run in SLI, does each card use less RAM, compared to running on a single card only? Do both cards load all texture in, or only those required to run on half of the screen?

 

What has happened to the good ole days? My Ti4200 and 8800 weep.

 

Assuming if Nvidia ever allows their AIB to release GTX970 8GB variant and game/bench in future is able to load that much texture/shading/etc into those memory, the impact would be more profound.

Since Nvidia officially mentions about the segregation of 3.5GB and 0.5GB for 4GB, we can speculate memory segregation for 8GB variant would be 7GB and 1GB. This allocation is quite inefficient in term of cost, power and design.

 

What about the partners, maybe they could be messing with those "extra" 500mb in the thought that people wont use them anyway. Could explain why only some users have the problems.

 

Some of you guys with single 970s purposely running out of gpu power with awfully high eye candy, aa and high resolutions have bad framerates when you load 4gb of vram? I'll alert the press.

 

Tbh, until some credible tech site does some thorough tests on the 970 to examine this, we cannot definitely conclude that the card is hobbled past the 3.5gb mark. Several individual users have tested the theory with conflicting results. Those that claim issues should submit their config, settings, etc for others to try it out. These same issues should hamper a 3gb 780ti (IF its a vram issue). 780ti's have been out there for almost 2 years with no one complaining, let alone SLI owners. It still beats any 4gb card out there (other than 980) in the most demanding situation.

 

Since I'm not sure if you're being serious or not....

Unless you have an AMD GPU, Mantle will not work. AMD has not yet made Mantle available to NVidia and Intel as it's still in "development".

 

Thread was reopened for comments about mantle?

 

And you are right, but I do not like words like "beats". If 780Ti runs into issue due to only 3GB of vram, users will lower few details. What I meant is: point where users will have to do that will come sooner to SLi 970 than if it had fully performing 4GB.

I stole this from post earlier:

When I saw it, I felt sadness. Dumping all details for sake of "Playing" @4k.
And still only have under 30 fps in static situation. And it looks worse than GTA 3. Rotating viewport may result in single digit fps and heavy stutter.
Anyone playing at 1080p would have game looking much better.
But it confirms that Single card of this design does not have enough brute force to process situations requiring 4GB vram anyway.

 

I think it was a big mistake for so many to have taken that cuda vram bench as a definitive sign of trouble for 970's gaming performance. It only highlighted a lowered bandwidth issue specific to the conditions of that program, not in actual gaming.

 

And for sake of testing:
I would like someone who understands fully what I write here to test.
(not following perfectly will result in worthless result)
- Have 2 games at hand:
- - Both must have windowed mode support. And have to be pre-configured to start in windowed mode.
- - 1st one which utilizes 3.3~3.5GB of vram.
- - 2nd which uses around 500MB of vram and preferably should have built in benchmark.
- Start 1st (big) game, let it load in windowed mode.
- minimize 1st game and keep checking that it sill uses 3.5GB vram
- start 2nd game and confirm that close to 4GB of vram are used
- - if so, run bench
- end both games
- start only 2nd (small) game and bench
- compare results and observe smoothness difference

It is necessary to be sure that 1st game resources are not unloaded. I tested it on my AMD card, which tells little about nV behavior. But if games are run in FS mode, I can run 10 at once and second I go off fullscreen (back to desktop and start next game) resources are unloaded. While running several games in windowed mode eats more vram with each game.

 

That's what the handful of people claiming that this "issue" was such a huge deal wanted. Their intention was to cause damage to NVidia by whatever means necessary. If that "benchmark" was run properly, it would show similar behavior for all cards. That's why they insisted that it be done with no memory load on the card. In real-world scenarios, all cards will produce similar results....which disproves their case and was thus not in their interest. -Tj- actually proved that early on in this thread.

 

Wouldn't CPU itself will present a issue?
Eg: If the first game already uses almost 4 threads? (Assume perfect cpu scaling for the game) Then second game would be bottlenecked by it? CPU does involved in issuing draw command to GPU.

 

Two games running at the same time https://www.youtube.com/watch?v=xMA9xKn0DaE

 

Hey guys today I was in the process of buying myself a Gigabyte GTX 970 G1 Gaming GPU. Based on this new problem, should I not buy it right now ? Plz advise. Thank you.

 

From my observations and comparison of 970 vs others shown here and on other sites.
970 shows with this bench degraded performance in last 512MB + memory taken by OS.
Others show degradation only on region taken by OS.
And should be noted that doe to chunk size of 128MB in default test, last chunk was never allocated therefore if OS taken less than 128MB card shows full performance till "end" which leaves last chunk (OS allocated) untested.

I call it an issue which has been overridden.
(In similar way as that rare Phenom bug which could cause freeze under heavy load. Back then override came at cost of 5 to 10% of performance based on workload.)

And notice, that test above 3.5 + 0.5GB allocation is worst case scenario and closely to impossible to get into in real world situation since in most cases nV will limit allocation to full 3.5GB.

I want to see properly done test and result just for sake of knowing.

 

Whether you buy now, or later makes no difference. The cards will be the same until they're discontinued. As long as you're running at reasonable settings for your hardware, there won't be any issues with performance.

 

No, moment you minimize it it will become docile. And main reason for 2 separate games using vram is:
- in most test difference is shown on already low performing games since they use 3.5GB of vram, ultra settings and AA.
- and using only one game where you change settings to increase vram utilization breaks entirely transparency of testing which is what I want most.

 

GJ upsetting the guy! He can't run 2 copies of Crew at the same time...