Anandtech UT2003 Benchies and the ATI 8500

[jb]

http://www.anandtech.com/video/showdoc.html?i=1608&p=5


Hmmm taken a closer look at the 8500 performance and it just does not add up.

800x600X32 = 40.9
1024x768x32 = 21.0
1280x1024x32 = 20.6


Now I dont want to claim a driver bug but does anybody here have a vaild reason why the 8500 would loose 1/2 its FPS score from 800x600 vrs 1024x768? Then why does it lose only .4 going to the next highest resolution? The 8500LE is in the same boat. Notice all the other cards dont show this effect. I have yet to see any other game (even the codemark creatutes scales well on the 8500) show this type of behavior. I just can not see it being fill rate or MB limited. And I doubt its hitting a TnL bottle neck. Its almost looking like its hitting another bottle neck that should not be there. Again .4 fps is just way to low going from 1024 to the 1200 res. In that type of case it should have fallen lower than the GF cards as the 8500 does not have that efficent bandwidth saving stuff.

So anyone have another expliantion other than a driver bug?

[Sharkfood]

So anyone have another expliantion other than a driver bug?

Yeah, *human error*. :)

[Doomtrooper]

So anyone have another expliantion other than a driver bug?

Possibly ATI-ISV can answer this ? :(

I'd really like to know what has changed in the builds that has taken the 8500 from one of the best peforming cards with UT 2 to almost the worst :x

http://www.anandtech.com/video/showdoc.html?i=1580&p=7

VS.

http://www.anandtech.com/video/showdoc.html?i=1608&p=5

[Reverend]

You can't really compare the two Anandtech articles (as per Doomtrooper's links above) - different UPT builds. Also, discounting all other cards, there's a difference in NVIDIA driver versions used.

I'm just guessing that it must be something to do with AGP or system memory allocation options.

[Mark N]

Look at Jedi Knight 2 in the same article - the LE does 99.8fps at 1024x768, slower than the GF4 Ti, slower than the GF3s, slower than the GF4 MXs. At 1280x1024 it does the same 99.8fps, now only the GF4 Ti is faster. At 800x600 the LE did only 99.3fps, and the 64mb Radeon was also slower than at the next step up. Maybe CPU limitations and the 128mb of memory explain all this, maybe not.

But regarding Unreal2, the Radeon has gone from 51.7fps in build 848 at 1024x768 to 21fps in build 918 here. The Ti 200 has dipped some as well, from 36 to 29.3, but that's ridiculous. My theory is that nVidia got to Sweeney and Co., laid out a little "incentive", and now suddenly the ATi hardware is fully "optimised". As the articles says, using this test "helps the guys at Epic during the development stage as they can work out driver bugs with the hardware vendors", or something along those lines...

[ben6]

Jedi Knight 2 : Jedi Outcast is an interesting benchmark. Until you enable FSAA and anisotropic filtering the benchmark doesn't in fact slow down much if at all in various resolutions . Take my runs on a Geforce4 TI4600 and Ti4400 on a 1800+ Athlon XP + nForce+256MB Ram sound disabled , High Quality settings (aniso turned off when appropiate in the graphics menu):

Ti4400 JK2 JO
============
1024x768
no FSAA 95.5
2X aniso 95.4
4x aniso 94.4
8x ansio 93.1


2x FSAA 95.2
2x aniso 94.8
4x aniso 93.1
8x aniso 91.2

4x FSAA 85.7
2x aniso 83.4
4x aniso 78.2
8x aniso 75.0

1280x1024 95.6
2x aniso 83.1
4x aniso 72.6
8x aniso 68.3

2x FSAA 91.9
2x aniso 76.6
4x aniso 67.4
8x aniso 63.3

4x FSAA 51.1
2x aniso 47.2
4x aniso 44.9
8x aniso 43.2


Ti4600
1024x768
no FSAA 95.6
2x aniso 95.6
4x aniso 95.0
8x aniso 94.7

2x FSAA 95.5
2x aniso 95.2
4x aniso 94.5
8x aniso 93.4

4x FSAA 92.5
2 x aniso 89.6
4x aniso 84.5
8x aniso 81.7

1280x1024
no FSAA 95.5
2x aniso 87.8
4x aniso 77.9
8x aniso 73.4

2x FSAA 94.8
2x aniso 83.6
4x aniso 73.9
8x aniso 69.6

4x FSAA 60.0
2x aniso 55.7
4x aniso 52.4
8x aniso 50.2

[Pete]

In the review, Anand mentions that ATi's scores dropped due to a driver problem:

Earlier we mentioned the performance anomalies at higher resolutions with the ATI cards and here you'll see exactly what we're talking about. Using the latest drivers just released last week the performance on all of the ATI cards is hurt pretty significantly here at higher resolutions. Even the GeForce4 MX 460 is able to pull ahead of the Radeon 8500 and the Radeon 8500LE. We'll be working with Epic and ATI to see if we can resolve these issues.

Luckily there are no visual artifacts by any of the cards/drivers.

[Reverend]

I see nothing in that quote that indicates it's a driver problem. Hopefully I'm wrong.

[Tagrineth]

I think in their first comparison they mentioned some flicker problems on the 8500.

[Doomtrooper]

I think in their first comparison they mentioned some flicker problems on the 8500.


Yes they mention a fog flicker issue that was resolved in the next test, which was the Geforce 4 review.

Link Here:
http://anandtech.com/video/showdoc.html?i=1583&p=12

In this review the tables seem to turn on the 8500, from the 1st build the Radeon 8500 is gettingc 51 fps @ 1024 x 768.
The link above the 8500 is actually faster @ 58 fps yet they state the 8500 is slower with the fog fixed...still confuses me on that comment.
This is also where all the Nvidia cards jump 30-40% in performance, be it the build or driver set they used it seems Unreal 2 is gonna run better on Nvidia hardware.
Hopefully one of the ATI guys can give us a answer why the 8500 is losing to a Geforce 4 MX :-?

[Windfire]

What I find funny is that Anand recommens the TI4200 128MB over the 64MB. Their reasoning is that you could potentially OC the 128MB version to match the memory speed of the 64MB version and that eventually "only" having 64MB will hurt you.

Hogwash. Under a few narrow circumstances. By the time you need more than 64MB for games 2-3 generations of video cards will have come and gone. Yes, the 128MB may be usefull if you want to run at 1600x1200 with FSAA but come on, lets get real.

I wonder if the reason why they didn't include the 64MB version was because it out-performs the 128MB version in 90% of the tests--which would tend to contradict their recomendation.

99% of the cards out there have 64MB or less and I would bet 95% of the cards to be purchased will continue to have 64MB or less.

I have yet to see a single real reason why having more than 64MB is so important as to go backwards in performance in everything we can play today (comparing the two flavors of the TI4200).

My guess is that in the next 12-18 months we might see one or two games that under certain "high-quality" settings with FSAA would like more than 64MB but by then we'll be drooling over much newer cards anyway.

Steve

[hughJ]

assuming bandwidth is still more of an issue than raw memory size, would there be some way of dividing up the 128 meg similar to that of how 3dfx did with vsa100, or is that really only beneficial when used with multichip SLI.. rather than say.. offering dedicated memory to a select number of pipes each.. it seemed amazing how easily one could output such high bandwidth numbers simply by scaling the ram like they did..

[Thowllly]

...
By the time you need more than 64MB for games 2-3 generations of video cards will have come and gone. Yes, the 128MB may be usefull if you want to run at 1600x1200 with FSAA but come on, lets get real.
...
My guess is that in the next 12-18 months we might see one or two games that under certain "high-quality" settings with FSAA would like more than 64MB but by then we'll be drooling over much newer cards anyway.

Have you tried Jedi Knight II: Jedi Outcast? On my gforce2 64mb card I can set all detail levels to max and it still runs smoothly, as long as I only use 16bit textures. If I switch to 32bit textures some parts of it becomes totally unplayable. And I've read that Soldier of Fortune II would also require a 128mb card for max texture sizes...

[Windfire]

Thowllly, yes, and on my 32MB Geforce2 GTS. Works great--and in 32bit color. Are you using texture compression? What resolution? With FSAA?

There are two choices:

1 - The Quake3 based game, with compression indeed needs more than 32MB, even more than 64MB video memory...

2 - There is something else going on here on your system

Option <2> is a lot more likely. I doubt any major game company would require more than 32MB in a game today with 32bit color. That is financial suicide.

We're back to the least-common-denominator here. With so many sub-32MB cards and with Geforce2 MX as the realistic target these days, companies that want to make $$$ shoot for full playability on mid-range systems.

[Pete]

People with 64MB Ti500's are reporting occasional stutters in certain RtCW maps, so 128MB may be prudent. The [H] JK2 benches are pretty persuasive, too. Though the 64MB may beat the 128MB in most benches, it isn't by a very significant margin. I'd gladly trade 10% in memory speed for no immense fps dips.

I see nothing in that quote that indicates it's a driver problem. Hopefully I'm wrong.
Yeah, I'm assuming it's yet another driver problem. It seems like almost every AnandTech review mentions the latestdrivers fixing something but breaking something else.

[Dave Baumann]

People with 64MB Ti500's are reporting occasional stutters in certain RtCW maps, so 128MB may be prudent.

I;ve just finished benching RtCW on the Ti 4200 (64MB) and there are some odd stutters - however it appears to get worse the second time a demo is loaded. It almost as though the textures are not getting cleared down from the first run and its trying to reload them all again (it takes a horrible amount of time to load the second time) - I'd say that this particluar issue is a driver glitch rather than a 64MB issue though.

[Windfire]

Pete, you bring up a good point. Frame rate stability is very important--one of the things I really liked about 3dfx when it was king--it may not have had as high a framerate but it was smooth all the time.

Heck, even with the slower RAM on the 128MB version, the board equals or surpasses the GF3 TI500 in terms of performance.

With this in mind, I don't think you could go wrong with either board.

[Sharkfood]

I'd be real interested to see some Serious Sam (FE and SE) ProfileStats with a 1 second interval graphed on the various cards (DumpProfileStats() if I remember correctly).

I'm undecided on the UT2 demo benchmarks. It's impossible to speculate what might be afoot with these benchmarks as Anand has no commentary or extra information. Most normal sources would question the massive variance, and silence I think speaks louder than such a comprehensive look at the scores. There is obviously something afoot, or a grievous error somewhere as those scores make absolutely no sense at all.

There is nothing cryptic, unreasonable or unsual about 3D graphics for the savvy. I believe the source/cause of this rather curious behavior would be fairly easy to identify by just about any other source- be it a driver bug, human error or some sort of coding issue in the demo that creates platform favoritism...

[Ailuros]

First of all can someone possibly explain why the UPT test hasn't been publisized yet? Or do the developers underestimate the average user as to not being able to run it by himself and reach a score and get a pretaste?

Speaking of pretaste: I fell for the scores at their very first appearance as to being indicative as to how UT2003 or any other U2 deriving game will react this year. At this point of time I doubt we will see in at least this year's games using that very same game engine that high poly counts as shown in the test.

If I'm even close in the above two assumptions then I consider releasing performance values from such a test not exactly as being so "innocent" after all, because it can eventually mislead the public by a lot in more than one occassions. Someone might say that Anand has pointed out that it's to be considered as an extreme stresstest along the lines, but it apparently wasn't made clear enough. And I don't like seeing benchmark numbers from it constantly reappearing when I as a user cannot in fact d/l and verify the results.

[jb]

First of all can someone possibly explain why the UPT test hasn't been publisized yet? Or do the developers underestimate the average user as to not being able to run it by himself and reach a score and get a pretaste?

Hi Ailuros,

I will take a stab at this and suggest that its because they have not finalized the engine yet and giving out the test now, could very welll make it harder to compare scores for cards today vrs a few months from now when the UT2003 is released to the public. I know they were trying to integrate some more stuff into their latest build (perhaps the Karma Physics code). Now while tweaking the engine here and there, it could lead to a difference in performance and thus provide no way to compare numbers. Thus I don't think publicly displaying numbers on non-final code is a good idea. Further more, I would want to see benchmarks from actually gameplay. Yes stress test are all fine and what not, but its not the same thing. I guess I would like to see both a in game and extreme test. But wait till its final.

BTW Ailuros, did you happen to see recently that a UT mod team released their long awaited update including a map that you help tested? hehehehehe

[Ailuros]

No I didn´t but I´m definitely happy for you. You might want to give it a shot in openGL with the latest renderer and "smooth alpha" enabled. Maybe you´ll get an idea why I found that a certain setting boosted my performance back then ;)

[jb]

Ok thanks for the tip good sir and check out the featured mod of week at PU to see it :) And beware of the proxies....

[LittlePenny]

It almost as though the textures are not getting cleared down from the first run and its trying to reload them all again

Very interesting thought. I don't have any recent games besides SimGolf, but maybe someone knows of a DX texture intensive game where we could test this? Maybe we could find that it is a OpenGL texture read/write error, or if it stutters again, I guess it could be either.

[RaolinDarksbane]

I don't understand why those games (SoFII or Jedi Knight Outcast) would require a 64+ MB videocard so they would stutter. Does those games uses S3TC? What's the point of texture compression if it isn't used properly?

[Randell]

maybe the issue is that everyone is so used to putting everything on max geometry & texture detail they dont realise that they should be putting newer games down a notch or 2. After all the average mid-1999 card (TNT2, Savage4, V3) didnt play Quake IIIA smooth at HQ settings when it first came out did they?

BTW I'm finding SoF2 MP test a lot smoother than MOH:AA and JKII and SOF2 test has very nice hi-res textures (they look like UT 2nd CD style or maybe just good detail texturing layering).

[Doomtrooper]

I can't say who my source is but I can tell you running the latest DEV drivers there is up to 50 % increase compared to Anands benches, and this was on a fairly lower end Athlon setup.
:P

[Joe DeFuria]

Do you mean the latest leaked dev drivers, or dev drivers that haven't yet been leaked to the public?

[Doomtrooper]

Do you mean the latest leaked dev drivers, or dev drivers that haven't yet been leaked to the public?

I mean the latest, latest fresh off the press drivers.not leaked yet :wink:

[Sabastian]

Do you mean the latest leaked dev drivers, or dev drivers that haven't yet been leaked to the public?

I mean the latest, latest fresh off the press drivers.not leaked yet :wink:

50% increase? OMG I guess the next question is how long before they are leaked? A 50% increase??? I can't believe that, are we talking about ATi here ? Are you sure? ;)

Sabastian

[Doomtrooper]

Do you mean the latest leaked dev drivers, or dev drivers that haven't yet been leaked to the public?

I mean the latest, latest fresh off the press drivers.not leaked yet :wink:

50% increase? OMG I guess the next question is how long before they are leaked? A 50% increase??? I can't believe that, are we talking about ATi here ? Are you sure? ;)

Sabastian

Notice Up to 50%....the information is coming from a reliable source..no FUD.

[Sharkfood]

That doesn't sound unbelievable.. 50%.

After all, we saw a 120-160% improvement in GLExcess scores once the mixed-mode texturing bug in OGL was fixed.

As we cant see the UT2-demo test, it's anyone's guess how applicable a specific driver bug.. or whatever it may be.. these scores reflect. It's pretty obvious from looking at the scores that something is seriously afoot with them.

It's also pretty obvious that this same kind of trend really hasn't been uncovered in *anything* else out there- so if it is indeed a driver bug, it would make sense it would not have been uncovered or isolated until now.

[Sabastian]

Notice Up to 50%....the information is coming from a reliable source..no FUD.

Sweet, I have some frends that will be extremely impressed with this bit of news. Thanks Doomtrooper. Will this be a specific API increase like Sharkfood pointed out about GLExcess scores or an overall increase? ATi is really changing impressions about their driver support....good news indeed.

Sabastian

[jvd]

http://www.anandtech.com/video/showdoc.html?i=1609&p=20


ati aiw 8500 128 gets 35.4 @ 1-24x768....

[Sharkfood]

Okay.. Now I'm *REALLY* confused.

Anand states:
"It would seem as if the poor scores in the previous roundup could be attributed to a combination of an older BIOS revision on the EPoX board and the latest ATI drivers."

First things first, on the driver spec page it lists the 6043 drivers. Then on the driver spec page of the previous 21-23 fps benchies, it also lists the 6043 drivers. Is there some typo or something I'm missing here? How could the same drivers "mature" between runs?

I'm happy Anand finally figured out how to flash a BIOS. This is good news and should be kind of "status quo" when testing systems. Somehow I doubt this had much impact on the improvement though.

Lastly, I'm confused about the 8500 tests (not the AIW 8500DV 128MB). The stock 8500 has changed from 21.x to 26.x fps between the tests. This is still far, far, far below what it should be reading (roughly half) when compared to the Ti500.

I have no idea what's up with this test but I really wish some site with some credence or PC/diagnostic/3D skills would pick up an 8500 and actually look into this with some hopes of creating some clarity on the matter. It just seems the more Anand posts, the more bizarre and strange the results.

[Reverend]

As much as I'd like to be in Anand's position where I continually get the latest UPT build, I would simply suggest that we discard Anand's UPT's scores, in the past as well as present. Too many stuff to cover and with Anand not (bothering?) to ask what sort of changes every new build means (whole lotta stuff to cover!), it's simply a guessing game. It would've been easier if Anand stuck with the same build for every tests/reviews but I suppose that's unlikely (even for me!).

Until, a detailed explanation of what the test involves (in terms of changes, from build to build) and this is updated for every build in every article, it really is a "see it, read it, forget it". For me, at least.

[McElvis]

In Anand's latest article, the ATI All-in-Wonder Radeon 8500 128MB, he re-ran the Unreal Performance Test 2002.

Here's what he states :

Unlike what we saw in our latest roundup, the Radeon cards do a lot better in the UPT benchmark. It would seem as if the poor scores in the previous roundup could be attributed to a combination of an older BIOS revision on the EPoX board and the latest ATI drivers. We'll be re-running those scores right away and will update the review accordingly.

So, a simple mistake?

[Sharkfood]

Read my first reply concerning "human error" :)

I'm with Rev on this one. All the UPT scores that have poured out of this site are pretty much useless information.

As there has been little to no emphasis on what's actually being tested from revision to revision, along with a trend that fails to identify or question unusual behavior, this kind of invalidates these scores from serving any purpose. The results are totally in a vacuum.

Add to this a disparity in results on nearly identical conditions between two reviews.. (only documented to be a simply BIOS flash)... it all adds up to some rather self-defeating, confusing and literally unusable data.

[Doomtrooper]

I will be honest, I think Anandtech has gone down hill bigtime, his CPU articles, motherboard etc... are usually quite good but his Video Card reviews have gone down hill.
If you look at the older 3DFX vs. Nvidia Vs. ATI comparisons there was lots of screen shots and much more indepth testing. :-?

[Sabastian]

Are these the ones that you are talking about Doomtropper? According to this thread they are old, relatively speaking of course.. ;)

http://www.rage3d.com/board/showthread.php?s=8a15cada38ac3670438fb57f7acb794e& threadid=33615048

[Daemon_UK]

Are these the ones that you are talking about Doomtropper? According to this thread they are old, relatively speaking of course.. ;)

http://www.rage3d.com/board/showthread.php?s=8a15cada38ac3670438fb57f7acb794e& threadid=33615048

Yes, enquiring minds would like to know! :evil:

[Sabastian]

Not that there is any problem or anything because there is a performance boost with these. But it isn't any 50% boost. ATi is pumping the drivers out in a big way. Looks like they really want to bury their poor driver reputation. I've been impressed anyhow.

Sabastian

[Doomtrooper]

I'm sorry guys my information came from a very good source and I must honor their request and keep my big yap shut :P
I just wanted 8500 owners to know (since this game will be my next jump in professional play) that the performance numbers are better than what Anand is posting.

8)