"I want to futureproof my videocard purchase and will choose a videocard with the most videocard memory" - Investigation

[RussianSensation]

UPDATED: September 13, 2008 - 4850 512mb vs. 4850 1GB

Call of Duty 4 - 2560x1600 8AA/16AF
20.7 vs 22.1

Clive Barker's Jericho - 1600x1200 8AA/16AF
28.2 vs 27.7
*Note at 2560x1600 4AA/16AF the 1GB card got 14.9 vs. 7.2 for the 512MB (but unplayable nonetheless)

F.E.A.R. - 2560x1600 8AA/16AF
33 vs 33

Gothic 3 - 2560x1600 noAA/16AF
19.0 vs 19.0

Rainbox Six: Vegas - 2560x1600 4AA/16AF
22.0 vs 21.5
*Note at 2560x1600 8AA/16AF the 1GB card got 19.9 vs. 6.1 for the 512MB (but unplayable nonetheless)

S.T.A.L.K.E.R. - 2560x1600 noAA/16AF
20.2 vs 20.4

Unreal Tournament 3 - 2560x1600 8AA/16AF
21.5 vs 29.0

Assassin's Creed - 2560x1600 4AA/16AF
28.8 vs 32.4

Bioshock - 2560x1600 noAA/16AF
26.4 vs 25.8

Call of Juarez - 2560x1600 4AA/16AF
12.4 vs 12.9
*Note at 2560x1600 8AA/16AF the 1GB card got 9.8 vs. 3.9 for the 512MB (but unplayable nonetheless)

Company of Heroes - 2560x1600 4AA/16AF
24.5 vs 25.7
*Note at 2560x1600 8AA/16AF the 1GB card got 23.9 vs. 16.9 for the 512MB (but unplayable nonetheless)

Crysis - 1920x1200 4AA/16AF
9.8 vs 10.0

Lost Planet - 2560x1600 4AA/16AF
15.7 vs 15.8
*Note at 2560x1600 8AA/16AF the 1GB card got 14.7 vs. 4.8 for the 512MB (but unplayable nonetheless)

World in Conflict - 2560x1600 4AA/16AF
27.5 vs 31.9

Conclusion: 4850 1GB provides no tangible benefit over 4850 512MB that warrants spending more $ at this current time, until new games arrive. But we can already see that when memory becomes a limitation, the card can't produce fast frames anyways. So 1GB seems more suitable for cards much much faster than 4850.



-----------------------------------------------------------------------------------------------------

September 9, 2008 - 8800GTS 320mb vs. 8800GTS 640mb vs. 3850 256mb

Q1: So how often do we hear that buying a videocard with more memory might be more "future proof"?

Q2: Is it worth spending $100 more for more memory all things being equal? How concerned should one be with choices such as 4870 512 vs. GTX 260 896mb?

Q3: Will the card of that "current" generation (when you purchased it) even benefit from that additional memory 6 or 12 months down the line? Or could be that some by the time that additional memory will be required, the card in question will be too slow anyway?

I stumbled upon an interesting review of modern cards by Tom's Hardware hoping to come out with some answers.

Disclaimer: Obviously ATI and NV use their memory differently in terms of efficiency but if memory is a concern, the 3850 256mb should easily fall to the 8800GTS 640MB card in the latest modern games (Unless shader power is more important than memory for that specific game). I also understand the comparison isn't entirely fair since some games benefit ATI (Assassin's Creed) while others benefit NV. But let's see if we can draw some conclusions regardless.

8800GTS 320mb vs. 8800GTS 640mb vs. 3850 256mb

Assassin's Creed- Max HQ
1280x1024 0AA - 41.5 vs 42.8 vs 54.8
1680x1050 0AA - 33.5 vs 33.9 vs 42.5
1920x1200 0AA - 29.0 vs 29.3 vs 34.8

1280x1024 4AA - 28.4 vs 31.1 vs 34.8
1680x1050 4AA - 24.9 vs 25.5 vs 26.3
1920x1200 4AA - 19.7 vs 22.2 vs 23.2
*There is no tangible benefit of having extra memory in AC since 8800GTS' gpu is too slow.

Call of Duty 4 - Max HQ
1280x1024 0AA - 82.2 vs 81.2 vs 83.2
1650x1050 0AA - 67.5 vs 70.2 vs 66.8
1920x1200 0AA - 57.0 vs 58.1 vs 55.5

1280x1024 4AA - 66.0 vs 65.2 vs 55.5
1650x1050 4AA - 49.4 vs 55.4 vs 43.0
1920x1200 4AA - 44.0 vs 47.2 vs 34.0
*While the 640mb card beat out the 320mb card, the performance advantage with AA is not large enough in any setting to provide much better playability .. in other words the gpu is just too slow.

Crysis - HQ
1280x1024 0AA - 25.6 vs 26.8 vs 25.5
1680x1050 0AA - 18.5 vs 20.5 vs 19.9
1920x1200 0AA - 14.4 vs 16.8 vs 10.9

1280x1024 4AA - 7.1 vs 18.5 vs 10.3
*640 produces faster frames than the 320mb card in 1 resolution, but once again its performance is too slow to matter in any resolution for a first person shooter game.

Crysis - Very HQ
1280x1024 0AA - 14.4 vs 17.1 vs 17.3
1680x1050 0AA - 11.5 vs 13.3 vs 12.6
1280 1024 4AA - 1.0 vs 11.5 vs 8.4
*640mb is 11.5x faster than 320mb card in 1 resolution...but nothing here is even remotely playable to begin with in any setting! The gpu is too slow.

Enemy Territory: Quake Wars - HQ
1280x1024 0AA - 80.1 vs 81.2 vs 69.7
1680x1050 0AA - 71.8 vs 76.8 vs 67.6
1920x1200 0AA - 61.5 vs 68.6 vs 63.7

1280x1024 4AA - 73.3 vs 71.8 vs 62.2
1680x1050 4AA - 60.0 vs 60.3 vs 50.4
1920x1200 4AA - 47.4 vs 49.8 vs 41.1
*Again we see that 640mb is marginally faster than the 320mb card without AA, but with AA enabled it hardly provides any more playability.

Half-Life 2: Episode 2 - Very HQ
1280x1024 0AA - 102.7 vs 103.6 vs 108.0
1680x1050 0AA - 95.2 vs 95.2 vs 98.3
1920x1200 0AA - 81.8 vs 81.9 vs 77.5

1280x1024 4AA - 96.1 vs 96.8 vs 68.7
1680x1050 4AA - 75.0 vs 75.8 vs 52.3
1920x1200 4AA - 59.6 vs 60.7 vs 41.4
*It would be easy to point out that 3850 is being limited by its memory here, but that would be the wrong conclusion. 3850 256mb CF gets 76.7 frames in 1920x1200 4AA (so the culprit is 3850's inferior AA performance off its shaders). Once again, there is no tangible benefit here for having extra memory. Of course this is an old game, so hardly a surprise.

Mass Effect - Ultra Q
1280x1024 0AA - 68.9 vs 69.4 vs 64.3
1680x1050 0AA - 54.1 vs 54.4 vs 49.1
1920x1200 0AA - 42.7 vs 43.0 vs 38.7

1280x1024 4AA - 48.7 vs 50.2 vs 27.3
1680x1050 4AA - 35.6 vs 36.7 vs 20.7
1920x1200 4AA - 17.0 vs 21.7 vs 13.3
*3850 shows again it's the gpu's AA performance that's the culprit here, not the 256mb of memory, as 3850 256mb CF gets 28.7 frames in 1920x1200 4AA.

1280x1024 8AA - 46.2 vs 47.5 vs 24.3
1680x1050 8AA - 31.3 vs 35.3 vs 15.3
1920x1200 8AA - 0.0 vs 21.2 vs 5.6

*Once again, we see that the 640mb card provides only a marginal improvement hardly worth mentioning, except in 1920x1200 8AA where it's already pretty slow.

Microsoft Flight Simulator X SP2 - Ultra Q
1280x1024 0AA - 32.8 vs 32.3 vs 24.0
1680x1050 0AA - 32.4 vs 32.9 vs 25.5
1920x1200 0AA - 33.3 vs 32.8 vs 25.0

1280x1024 4AA - 25.1 vs 34.4 vs 26.0
1680x1050 4AA - 3.0 vs 29.3 vs 22.1
The 640mb card definately provides a serious advantage over the 320mb card here. Considering ATI's card only has 256mb of ram, this is an illustration of NV's poor memory management.

World in Conflict - Very HQ
1280x1024 0AA - 18.0 vs 32.0 vs 28.5
1680x1050 0AA - 16.7 vs 28.8 vs 27.9
1920x1200 0AA - 14.2 vs 24.8 vs 24.3

1280x1024 4AA - 9.5 vs 22.0 vs 15.2
The 640mb card clearly has an advantage here. Again we see NV being outperformed by the 256mb card due to its superior memory management.

Conclusions

1. Yes there will be situations where extra videocard memory will actually provide superior performance (it would also be interesting to investigate the impact on minimum frames and not just averages), but...

2. ...In a lot of cases the card's GPU will be a more serious bottleneck.

3. Considering one could have saved $100 when buying the 320mb version of the card, today that $100 + $50 would buy 4850 which provides significantly faster performance than 640mb. At the time of purchase though, the 320mb card was not slower than the 640mb card. Today, the 640mb is hardly fast enough even in situations where it has the added memory advantage.

We have seen this already many times before - just to name a few:

8500 Pro 64mb vs. 128mb (hardly worth it)
9800 Pro 128mb vs. 256mb (hardly worth it)
X850XT 256mb vs. X850XT 512mb (hardly worth it)
X1800XT 256mb vs. 512mb (the 512mb didn't really provide "more life" as in 6 months it was too slow anyway)
8800GT 512mb vs. 8800GT 1GB (no difference today and in 12 months it'll be too slow anyway)
etc.

4. (Something to keep in mind: It might not be correct to say that ATI xMB = NV xMB after we have seen 256mb 3850 outperform 8800GTS 320 in several benches. Although, the memory mismanagement might be specific to the 320mb card, so I won't draw this conclusion).


So what does this mean for your next videocard purchase?

1. If the price difference between the 2 cards in question isn't significant, it probably makes sense to get the card with the larger memory.

2. However, more likely than not in 12 months from now with more intensive games we'll hardly see GTX 260 896mb provide a significantly faster performance than 4870 512mb OR 4870 1GB outperforming 4870 512mb. More likely than not it is their GPUs which will start to become significantly slower, negating any memory advantage.

3. So if you are having a hard time deciding between a 512mb vs. 896mb vs. 1GB version of a card with a similar gpu speed and memory bandwidth, you'll probably be better off not spending extra $ and saving it towards an upgrade in 12 months anyway. In this case, focus on other features instead (noise/power consumption/image quality/HD acceleration/physX support, etc.).

Just some food for thought.


P.S. Anyone who is able to offer benches of 512 vs. 1GB of the same cards, please feel free to add to this thread!!

EDIT: I will try to find more benchmarks tonight for 4870 1GB vs. 4870 512mb or other relevant ones to add to the discussion to expand it.

 

[cmdrdredd]

What I figured...it also depends on the resolution and things like that.

 

[apoppin]

i am going to contradict the OP and Toms'

Is the 512MB 4870 insufficient for 19x12 ?

Don't buy a 4870/512MB if you want to game at 19x12
- it is NOT smooth even when the FPS stay ABOVE 40 FPS in some games!!!
:Q

My Sapphire HD4870/512MB is a *dog* at 19x12 - even though the FPS *appear* to be playable
:Q


my conclusion so far, and the original reason i was SO disappointed with mine - especially in comparison to my 8800GTX or 2900xt Crossfire

more background about my comparisons:
http://alienbabeltech.com/?p=270

i may upload a FRAPS video - especially for ET:QW and Crysis - it is noticeable. Unfortunately just recording a FRAPS video kills the frames and it still takes forever to upload on Satellite Broadband

 

[RussianSensation]

apoppin, the only way to actually say that 512mb is insufficient for 1920x1200 on 4870 is to test the 1GB card to compare it to. If the 1GB provides much smoother performance, then we can make this conclusion. I also realize there could be significant drops in performance for minimum frames despite similar averages (sudden dip to 3-5fps). I think it should be a requirement for videocard websites to show average AND minimum frames.

It all rests on you buddy

 

[apoppin]

Originally posted by: RussianSensation
apoppin, the only way to actually say that 512mb is insufficient for 1920x1200 on 4870 is to test the 1GB card and show that it is sufficient in those games with same settings.

It all rests on you buddy

Look, if you run a 2900xt, a 8800GTX, a HD4870x2 - and it's minimum FPS stays above 30 - no matter how much it is loaded down with filtering and maxed details - it does not start to get "choppy" like my HD4870 does when it runs above 40 FPS [sometimes, especially in the games i mention]

it .. stutters .. it is NOT SMOOTH for the FPS it is running
-Does anyone else notice this?

All i have to do it turn down the resolution and it becomes much less noticeable

it is really obvious in ETQW and Crysis .. not so much in UT3 - even at 20 FPS in UT3 it is not like in the other games at 30FPS

Framebuffer is full - what else could it be?
--do i have a defective card? i doubt it
i will go ahead and re run those ETQW benches again with HD4870x2 with a core disabled... But i need to finish benching with 4870 first and then run 2900xt .. first ..

 

[VirtualLarry]

Oh noes! Does this mean that my purchase of a 512MB GeForce 6200LE was foolish?!?

 

[nyker96]

I can't dispute these real world results but I have seen some big differences in benchmarks between 8800GT 256 vs 512 battle. Perhaps for some generations of GPUs the higher memory will show through.

 

[RussianSensation]

Originally posted by: apoppin

Framebuffer is full - what else could it be?
--do i have a defective card? i doubt it

You bring an interesting point - we need a monitoring software for videocard framebuffer access. What do we have for this? RivaTuner!!!

Main --> Customize (top menu under Target Adapter) --> Hardware Monitoring (magnifying glass over the memory chip logo) --> Setup (bottom right corner) --> Plugins (bottom left corner):

Select 1) Videomemory usage
2) Local videomemory usage
3) Non-local videomemory usage

Voila!! Now you don't have to 'guess'.



P.S. Please post videomemory usage pics when comparing your 1920x1200 benches.

 

[RussianSensation]

Originally posted by: nyker96
I can't dispute these real world results but I have seen some big differences in benchmarks between 8800GT 256 vs 512 battle. Perhaps for some generations of GPUs the higher memory will show through.

That's true. I think the point is 8800GT was a mid-high end card and at the time 512mb was mid-high end. So you want to try to match the videocard memory to the type of card you have. Right now 512mb seems to be the sweetspot (unless Apoppin shows tangible differences )

As VirtualLarry pointed it there is no benefit in having 512mb of ram on a crappy slow card. In a similar fashion you don't want to nuter a mid-range card with 256 mb of ram.

But do we need more than 512mb right now? Doesn't seem that way.

 

[NoSoup4You]

Apoppin is spot on. I noticed this several years back when I had two 7900GT 256MB's in SLI with a variety of games. The fps would be nice and high (45+) at all times but the game was sluggish as can be. This was with my current rig in sig, all components remain the same (dual core cpu, 2 gig's RAM...etc.).

The resolution I've gamed at for almost 7 years now is 1600x1200. Not the hugest resolution by today's standards, but still a huge leap from 1280x1024.

Oblivion was one of the biggest culprits. After the user mods started piling up (higher resolution texture mods), the game started to become choppy even though FRAPS read my fps as being over 45 fps at all times. Turning my character would introduce 'jitters', it was sluggish despite a rock solid framefrate. This has to be something related to the memory, one would think.

HardOCP has discussed this in a few articles. As I recall, their review of 8800GTX's in SLI at 2560x1600 backed up this observation. They were seeing frames higher than 40 and yet the game felt too sluggish to be playable. Dropping the AA or solved it.

It might be hard to understand how 45+ fps in a game can feel sluggish until you see it in person.

 

[NoSoup4You]

I dug up the article. http://enthusiast.hardocp.com/...l?art=MTI2MiwxMCwsMA==

The part I'm referring to -

Do we need more graphics memory?

In this evaluation, especially under Oblivion and some other games, we talked about some odd performance experiences. We said that at certain very high settings we experienced a ?choppy? or ?laggy? feeling in the game even though the framerate indicated playable framerates. It may be that the games are able to render fast enough with 8800 GTX SLI but it may not be able to dump that information to video card memory fast enough, or, it may be running out of framebuffer space on the video card. Could it be that 768 MB of RAM on the video card is simply not enough for certain settings like 2560x1600 4X AA in Oblivion or motion blur being enabled in Need for Speed Carbon?

It may very well not be enough considering how much demand is on the memory at those settings. Perhaps 1 GB or more of graphics memory on video cards doesn?t sound so unorthodox now. Certainly it may not benefit single-GPU video cards as much right now, but with two 8800 GTXs in SLI running at 2560x1600 1 GB or more of graphics card RAM per video card, with more memory bandwidth may just be the ticket for letting the 8800 GTX GPU in SLI really shine. We saw the framerates it can render, but we just can?t play at those levels yet with the current bottlenecks, whatever they may be.

 

[apoppin]

OK, when i enable the plugins, all i see are "Page file usage" and "physical memory usage" in the actual HW monitoring screen

 

[myocardia]

Originally posted by: VirtualLarry
Oh noes! Does this mean that my purchase of a 512MB GeForce 6200LE was foolish?!?

No way, man. That card will still tear through Crysis @ 2560x1600/Very High. How could it not? It's got 512MB of VRAM, after all.

Originally posted by: nyker96
I can't dispute these real world results but I have seen some big differences in benchmarks between 8800GT 256 vs 512 battle. Perhaps for some generations of GPUs the higher memory will show through.

Yeah, I've seen those benchmaks, also. It just goes to show that you don't ever want to buy a card with less VRAM than all the other cards are sporting, even if it's got the GTX500000000000000000000000000000 GPU on it.

 

[apoppin]

Originally posted by: apoppin
OK, when i enable the plugins, all i see are "Page file usage" and "physical memory usage" in the actual HW monitoring screen

Bad news

- Videomemory usage monitoring is not available under Vista due to Vista videomemory virtualization.

NP, since RivaTuner can't show the usage, i can use my 4870x2 with a core disabled on the same scenes the 512MB version is struggling with - at 40 FPS

 

[angry hampster]

IMO the OP's findings are quite flawed due to the type of cards chosen. There should have been an 8800GTS 320/640 findings section, an 8800GT 256/512 section, and (if it's available) a 4870 512/1024 section. Including newer, substantially more advanced architecture that runs at higher frequencies with different memory modules is a terrible way to to show side-by-side comparisons. There's no point in having so many variables.

edit: here's an article with the 8800GT 256 and 512mb versions both reviewed. The 512mb version is ahead in every test.

http://www.legitreviews.com/article/607/7/

 

[dug777]

Originally posted by: angry hampster
IMO the OP's findings are quite flawed due to the type of cards chosen. There should have been an 8800GTS 320/640 findings section, an 8800GT 256/512 section, and (if it's available) a 4870 512/1024 section. Including newer, substantially more advanced architecture that runs at higher frequencies with different memory modules is a terrible way to to show side-by-side comparisons. There's no point in having so many variables.

edit: here's an article with the 8800GT 256 and 512mb versions both reviewed. The 512mb version is ahead in every test.

http://www.legitreviews.com/article/607/7/

He compared the same card with different ram Ignore the ATI card if it offends you that badly

Quite clearly, the extra RAM was not a good investment, then or now.

 

[BFG10K]

Originally posted by: apoppin

Look, if you run a 2900xt, a 8800GTX, a HD4870x2 - and it's minimum FPS stays above 30 - no matter how much it is loaded down with filtering and maxed details - it does not start to get "choppy" like my HD4870 does when it runs above 40 FPS [sometimes, especially in the games i mention]

If the problem was video memory then the 2900 XT would have the same problem, unless you have the 1 GB version of course.

 

[Lonyo]

Ignoring Apoppin's specific situation and what may or may not be responsible for it, he does make a good point and it may be his issue.

If you get say a 5% drop in average performance, but that comes due to a significant change in the way the frames come in (i.e more hits down to very low minimum FPS and more stuttering as stuff gets moved into the RAM) then the playability issue changes from "is the average FPS high enough to be playable" to "is the gameplay smooth or does it stutter sometimes?".
Yes, generally you're not likely to benefit a huge amount from extra RAM (resolution dependent) and while average numbers are useful, this whole arena is one where HOcp's benchmarking comes to be very useful when they show you a graph of frame rate over time so you can see how the average FPS is distributed throughout a given test.

 

[Keysplayr]

Originally posted by: BFG10K

Originally posted by: apoppin

Look, if you run a 2900xt, a 8800GTX, a HD4870x2 - and it's minimum FPS stays above 30 - no matter how much it is loaded down with filtering and maxed details - it does not start to get "choppy" like my HD4870 does when it runs above 40 FPS [sometimes, especially in the games i mention]

If the problem was video memory then the 2900 XT would have the same problem, unless you have the 1 GB version of course.

The 2900XT had a 512bit bus. Did it not? It also had a ring bus memory controller, which may have been more efficient managing the memory access. Maybe it's not the amount of memory we should be looking at here, but the bus width? Just an observation. Sure, the 4870X2 has a 256bit bus, but for each core yes?

I have to say, that I did notice a few games (Crysis being the most prevalent) where a 8800GTS 640 played smoother than a 9800GTX. Not faster framerates mind you, the 9800GTX blew the 8800GTS640 away in fps, but the disk thrashing was much more frequent with the 9800GTX. And it could be the slimmer bus on the 9800GTX could have been a contributing factor to over all performance.

Too many questions to be answered here that are not answered in the OP.
Nice Job BTW RussianSensation, but there are a lot of variables to eliminate.

 

[RussianSensation]

Originally posted by: apoppin

Originally posted by: apoppin
OK, when i enable the plugins, all i see are "Page file usage" and "physical memory usage" in the actual HW monitoring screen

Bad news

You followed the steps and you used a later version like 2.09 or 2.10? It for sure works. You should be able to use the scroll bar down which is why you didn't notice those selections?

 

[taltamir]

That 100$ saved then +50$ is not as impressive as saying....
Today that 100$ would buy a 9800GT which WILL outperform the the 8800GTS 640MB at no extra cost!

Yea, extra ram is nice, but not 100$ nice.

 

[Genx87]

My rule of thumb has been to get the 2nd best card in a generation. I got a 6800GT when the NV40 showed up. I got an 8800GTS 640 when the NV80 showed up.

Both cards worked well through the entire life of the gaming machine until the whole thing was replaced. And I didnt have to spend the extra cash on the high end part. Saving myself about 400 bucks between both cards.

 

[myocardia]

Originally posted by: angry hampster
edit: here's an article with the 8800GT 256 and 512mb versions both reviewed. The 512mb version is ahead in every test.

http://www.legitreviews.com/article/607/7/

You linked to the wrong page of that article, man. Try this page:

Stalker @ 1600x1200

256MB 8800GT Alpha Dog- 37.72 fps
512MB 8800GT @ stock clocks- 97.36 fps

 

[dug777]

http://www.anandtech.com/video/showdoc.aspx?i=3175&p=5

More of the same 8800GT comparo.

 

[apoppin]

Originally posted by: RussianSensation

Originally posted by: apoppin

Originally posted by: apoppin
OK, when i enable the plugins, all i see are "Page file usage" and "physical memory usage" in the actual HW monitoring screen

Bad news

You followed the steps and you used a later version like 2.09 or 2.10? It for sure works. You should be able to use the scroll bar down which is why you didn't notice those selections?

Yes the latest 2.10

it for sure doesn't work
- i C&P the message that it does NOT work in Vista
---this is from RT:

- Videomemory usage monitoring is not available under Vista due to Vista videomemory virtualization.

i am sure as hell am not going to set up XP just to monitor vRAM usage. it is easier to use my X2 with a core disabled

if it doesn't stutter with the X2, then the problem is with the framebuffer
[logically, by deduction; everything else is the same or slightly inferior on the X2]

If the problem was video memory then the 2900 XT would have the same problem, unless you have the 1 GB version of course.

not necessarily

it has a 512 bit bus and a slightly different architecture [ring bus]
- you also need to realize that i never used 19x12 to test it when i owned it
[however, i AM going to test it tonight over the same benches as the 4870 - so i will let you know if it has the same issues with hitching; i never noticed it before - not like with 4870]