3dmark05, the 'drivers' situation

[myshkinbob]

I thought i'd post this in the drivers section, as it's really about drivers and not 3dmark05 per se. Everyone's seen the latest catalysts seriously improved upon 3dmark05 scores, but not especially 3dmark03. So move this thread if you like, after you've read it, if you feel this isn't the correct place to post it.

This is a bit of supposition, so go along with me for the moment, but i think the ati gains aren't cheats at all, and that it really is a 'memory bug fix' of sorts. I may be wrong here btw, but, the cat official 4.9's on the PEG card, pci-express, gave the higher scores before the hotfix was even released.. Now pci-e is a different kind of host bridge, and it's entirely likely the memory interface and therefore on-card memory management is different, no fast writes, or agp aperature etc etc, so the card's memory management would be programmed differently to the agp cards, and that's how ati came across this 'bug fix'.

But it's a curious bug that once fixed doesn't show up any gain in '03 but does in '05. However, if you look at the full test suite for the pro version, there's something called batch size tests. I won't paste from the help file, you can read that. The gist is the IHVs don't optimise their drivers for different batch sizes, they try to push developers into using large batch sizes as much as they can. Now compare the batch size results for the x800 pro to the 6800UE, the pro romps it home. I couldn't find any results for an x800 card without the hot fix on this test, but i'm willing to bet the older driver doesn't handle it quite so well. The nvidia driver certainly doesn't.

So it's my rather wild and technically ill-researched guess that the memory fix improves batch size scores, and that the 3dmark team having stated good reason to make the IHVs improve that aspect of drivers, have allowed such improvements to have a bearing on the game tests also. It may be that the default memory block (and so batch) fetch size for the graphics card ram is rather large, and when the card is fed smaller block fetches than it or it's driver prefers, as in the batch fetch size tests, a lot of inefficiency occurs and cycles are wasted. If you could improve the batch fetch size efficiency aspect of the pipeline somehow, in your driver say, then any benchmark which isn't written to use the 'usual IHV pushed large batch size methods' would show marked improvement over previous drivers.

So that's my thoughts on it, prove me wrong, or prove me right, my poor technical knowledge and a lack of a pro account means i can't find the right technical details to prove or disprove and/or some non-8.07 ati batch size results.

Oh and feel free to move the post if this is more benchmarking than drivers. :)

[aAv7]

well, seeing as this is pertaining to ati drivers...i'm guessing this should be the the "other video card" section....

To me, the hotfix isnt a huge deal. I dont buy the memory fix, because the hotfix only helps out the r420 series, and memory problems have been known since the 9800 series.

To me, this driver is tweaked to death, by ati, to ensure they win this benchmark at launch and not receive another loss to nvidia. In 3 months, if nvidia is winning in this benchmark, it wont make as much noise...people wont really notice or care...but right now, during this benchmarks launch, the world is looking at who will win. Ati knew they needed this win. They lost big in Doom 3 and HL2 is still nowhere to be found....they needed something to show their cards are worth the $$. Now I dont feel they cheated...not at all, but I do believe they took an extra amount of time out on their drivers, for this fix and this benchmark, whereas nvidia is the type to wait for results, and then fix something.

The biggest reason I feel this way is because this hotfix does nothing for in-game performance. I've been to rage3d...some say there's no gain in games, other say roughly a frame....which means this driver was obviously put out for this benchmark. It's smart buisness...ati knew they needed this win, and they got it. I still feel nvidia needs to shape up on it's vertex shading, x800 pro's are more powerful than 6800UEE in that area, which is simply unacceptable. But other than that, I think nvidia is pretty much even with ati. My gripe is with the b.s of sm3.0...but that's a whole nother' thread. ( which i think i will start)

[Bumrush]

Well, either way people need to remember that 3dmark 2005 is not a game that you can actually play. As far as I'm concerned both cards are great when running games, you can't go wrong with either.

[subbo]

It might not be a game, but games like SW Battlefront use many of the same tricks and effects, and 2005 games like Battlefield2 look surprisingly similar in screenshots, in terms of self-shadows, bumpmapping and lighting tech. Games are permanently out of the Quake3 era after titles like Doom3 and HL2.

[kyleb]

To me, the hotfix isnt a huge deal. I dont buy the memory fix, because the hotfix only helps out the r420 series, and memory problems have been known since the 9800 series.

To me, this driver is tweaked to death, by ati, to ensure they win this benchmark at launch and not receive another loss to nvidia.

So why do the pci-express x800s still woop ass with the older drivers?

[Clay]

*** moved ***

[mustrum]

well, seeing as this is pertaining to ati drivers...i'm guessing this should be the the "other video card" section....

To me, the hotfix isnt a huge deal. I dont buy the memory fix, because the hotfix only helps out the r420 series, and memory problems have been known since the 9800 series.

To me, this driver is tweaked to death, by ati, to ensure they win this benchmark at launch and not receive another loss to nvidia. In 3 months, if nvidia is winning in this benchmark, it wont make as much noise...people wont really notice or care...but right now, during this benchmarks launch, the world is looking at who will win. Ati knew they needed this win. They lost big in Doom 3 and HL2 is still nowhere to be found....they needed something to show their cards are worth the $$. Now I dont feel they cheated...not at all, but I do believe they took an extra amount of time out on their drivers, for this fix and this benchmark, whereas nvidia is the type to wait for results, and then fix something.

The biggest reason I feel this way is because this hotfix does nothing for in-game performance. I've been to rage3d...some say there's no gain in games, other say roughly a frame....which means this driver was obviously put out for this benchmark. It's smart buisness...ati knew they needed this win, and they got it. I still feel nvidia needs to shape up on it's vertex shading, x800 pro's are more powerful than 6800UEE in that area, which is simply unacceptable. But other than that, I think nvidia is pretty much even with ati. My gripe is with the b.s of sm3.0...but that's a whole nother' thread. ( which i think i will start)

If it's a cheat, why did the x800 XT reach the same astronomic scores with the untweaked cat 4.9 then?
Your argumentation is lacking.
Nv got owned in 3dmark05 - no big deal it's only a benchmark.

[ChrisRay]

If it's a cheat, why did the x800 XT reach the same astronomic scores with the untweaked cat 4.9 then?
Your argumentation is lacking.
Nv got owned in 3dmark05 - no big deal it's only a benchmark.


Not entirely sure of that, This optimisation could very well be just a controller optimisation for the way the Graphic Adapter handles Vertexes over 256 megs (Which would frag system Memory)

Assuming this is the case. nvidia could implement a similar tweak. Its entirely possible.

[myshkinbob]

I wrote the post when i was a bit tired last night, the point wasn't meant to be about ati drivers cheating or not. What i was driving at is that the techniques used in 3dmark05, iterative resampling of shadow data in the graphics ram, must be atypical of the usual large fetch sizes implemented by developers at the behest of the IHV's. I'd presume optimising for different sizes of fetch increases the memory efficiency somehow. If futuremark made it seperate a feature test, you can bet it has an effect on the main benchmark, just as vertex/pixel shading performance, and fill rate, does.

Why i put the post in the forceware drivers section, is because what i'd like to know is, does the nvidia architecture control memory access in the driver, or is it hardwired into the graphics core itself. Now this is in the other desktop graphics card section, i don't suppose it's going to get much attention from anyone with a deeper knowledge of nvidia's architecture for the 6800 series.

Oh and those saying it's just a benchmark, it's not, it's a benchmark based on the techniques developers fully intend to use in the next year. For example the unreal 3 engine and carmack's next engine both use iterative shadow sampling to create soft shadows for all objects in the engines.

[DMA]

Nv got owned in 3dmark05 - no big deal it's only a benchmark.

Exactly.
I can't stand the nerds who put the score in their sigs just to brag.


Ooopsie. :D :)

[Riptide]

Any idea when we'll be seeing an official catalyst release of this new driver? I had some issues I thought might be tied to 4.9 so I'm still back at 4.8 for the moment.

[myshkinbob]

Exactly.
I can't stand the nerds who put the score in their sigs just to brag.


Ooopsie. :D :)

It's not about scores, it's about highlighting the different architectures performance in the forthcoming graphics techniques. It's sad that nobody wants to discuss why there's such a big difference between the two vendors in what is right now the only available next-gen 3d engine. I'm also pretty dissapointed at the moderating here right now, you could've at least moved the thread to the benchmarking forum, if you didn't see the relevance to the forceware driver memory performance. I mean, the other desktop cards forum is pretty much exile, very few people read and post here. I'm sorry the thread got dragged down to the level of vendor X cheats and benchmarking is for bragging rights.

[schuey74]

Any idea when we'll be seeing an official catalyst release of this new driver? I had some issues I thought might be tied to 4.9 so I'm still back at 4.8 for the moment.
These betas are going to be the 4.10s and considering that futuremark already approved them, I'd guess that they are on schedule for release next week, or the week after at the latest.

[jimmyjames123]

I still feel nvidia needs to shape up on it's vertex shading, x800 pro's are more powerful than 6800UEE in that area, which is simply unacceptable.

The vertex shader speed test essentially directly correlates to # of vertex shader units * core clock speed. The 6800GT/6800U/X700XT/X800Pro/X800XT all have 6 vertex shader units. The reason these Radeon cards have higher synthetic vertex shader speed than these Geforce cards is because of core clock speed differences. In reality, it seems that most current and possibly even future games are more limited by pixel shader speed than by vertex shader speed.

[mustrum]

Exactly.
I can't stand the nerds who put the score in their sigs just to brag.


Ooopsie. :D :)
If it's only about bragging then the whole nvnews forum is bragging.
I put it there for easy comparison. I like it when others do so as well.
If someone has a quite similar system as i do then i can click on it and compare them. This can help optimising the system in some cases or at least show if something is wrong.
Benchmarks not neccessarily show game performance but they are great to compare quite similar systems.

[nVestor]

If it's only about bragging then the whole nvnews forum is bragging.
I put it there for easy comparison. I like it when others do so as well.
If someone has a quite similar system as i do then i can click on it and compare them. This can help optimising the system in some cases or at least show if something is wrong.
Benchmarks not neccessarily show game performance but they are great to compare quite similar systems.

How is that? I mean, you have novice that buy a top-of-the-line OEM from various companies who don't know a thing about overclocking or system tweaking, while having someone with a less powerful system beating their score.

Then you have the guy who knows all the tricks about past Futuremark benching, runs one bench at a time, allowing the system to cool down before moving onto the next test, then combines them together to submit a complete bench. Meanwhile, during the bench they could be shooting a can of air into the HSF of the CPU or on the videocard to cool it down significantly.

These are the kinda things you use to validate a system? Kinda rediculous to do imho, given those kinds of variable.

Fun tool to goof around with, but I wouldn't take the scores seriously nor use it to compare similarity in system configs.

[myshkinbob]

Well i thought i should update with some info, firstly, the x800 core on any driver is significantly better on the batch size tests than the 6800s. Secondly, the memory bug fix did indeed improve memory bandwidth on the x800 line by about 10% according to the 3dm05 fill rate tests. So my batch size theory is indeed wrong, who knows where exatcly ati pulled the performance from. Also some 3dmark05 reviews have done IQ testing on the latest catalyst and concur there are no cheats causing a loss of IQ.

[DMA]

If it's only about bragging then the whole nvnews forum is bragging.
I put it there for easy comparison. I like it when others do so as well.
If someone has a quite similar system as i do then i can click on it and compare them. This can help optimising the system in some cases or at least show if something is wrong.
Benchmarks not neccessarily show game performance but they are great to compare quite similar systems.

I was only kidding m8. I used to have two compare URL's in my sig myself. :)
Have spent ALOT of time benching, 3DMark-01 in particular.
Now i don't really have the passion anymore, so time to play some games instead. :o

3DMark-05 is fantastic though, so i'll probably make a strong comeback this winter. :p

[Fahim]

This driver is a memory allocation fix, not a cheat. Performance improvements happened even with 9600XT 256MB in various games like Far Cry, Doom3, TRAOD. Also performance improves in AquaMark. All PCI-e X800 Series are getting the same score with CAT 04.9, because in PCI-e serieses memory allocation are already fixed. As AGP serieses are not fixed, thats why this driver.

[mustrum]

How is that? I mean, you have novice that buy a top-of-the-line OEM from various companies who don't know a thing about overclocking or system tweaking, while having someone with a less powerful system beating their score.

Then you have the guy who knows all the tricks about past Futuremark benching, runs one bench at a time, allowing the system to cool down before moving onto the next test, then combines them together to submit a complete bench. Meanwhile, during the bench they could be shooting a can of air into the HSF of the CPU or on the videocard to cool it down significantly.

These are the kinda things you use to validate a system? Kinda rediculous to do imho, given those kinds of variable.

Fun tool to goof around with, but I wouldn't take the scores seriously nor use it to compare similarity in system configs.

I hardly doubt that many people do that. I am sure most scores are legit. Sure some people will push the hardware to it's max wich they don't do when playing games but that's it. I am not talking about thos top 10 scores btw.
Those guys fighting for the lead will do everything to get a better score of course.

[neliz]

Those guys fighting for the lead will do everything to get a better score of course.

It won't be more than a couple of months before the 10K barrier is busted, this mornings scores saw the XT-PE reach into 8000's
The first 6800UE's (I don't know if it's SLI) marked their first 7000 score.

now tying the 800mhz XT-PE with that 6Ghz P4 and you should be able to get a nice overall score for 3dmark...

[mustrum]

It won't be more than a couple of months before the 10K barrier is busted, this mornings scores saw the XT-PE reach into 8000's
The first 6800UE's (I don't know if it's SLI) marked their first 7000 score.

now tying the 800mhz XT-PE with that 6Ghz P4 and you should be able to get a nice overall score for 3dmark...
Yeah but thanks got 3dmark05 shows the actual clocks of the graphics card again. If you see 700+ mhz on a XT PE you can ignore that benach since it's not a clock one could use for everyday use.
650mhz are possible for 24/7 use with a good XT PE and good cooling though.
(voltmodded of course but looking at the GTs... they mostly are voltmodded as well)

[Assi9]

The biggest reason I feel this way is because this hotfix does nothing for in-game performance. I've been to rage3d...some say there's no gain in games, other say roughly a frame....which means this driver was obviously put out for this benchmark. It's smart buisness...ati knew they needed this win, and they got it. I still feel nvidia needs to shape up on it's vertex shading, x800 pro's are more powerful than 6800UEE in that area, which is simply unacceptable. But other than that, I think nvidia is pretty much even with ati. My gripe is with the b.s of sm3.0...but that's a whole nother' thread. ( which i think i will start)

You might want to read the forums again. Many people have experienced increases in Doom3 and other OpenGL games as well as UT2K4, and to some extent; Farcry. I for one went from 40fps to 47fps in Doom3(1280x1024 4AA/16AF -- Demo1) from these drivers, and havn't yet enabled AI. Furthermore I don't get as many hitches in the game with these new drivers.

Since this is me tho, I can't speak for everyone else.

[Dr_s99]

I don't get it....

go read this article... http://www.megagames.com/news/html/hardware/atinewdriversget3dmark05crown.shtml
which is called "ATI New Drivers Get 3DMark05 Crown", in megagames.com
and tell me if you find it stupid....

since they say that X800 xt pe, gets higher score then SLI of nvidia....

but all of the sudden the say that X800 xt pe is overclocked....
now i don't get how this related to the driver?

if it does relate,
so does this mean that ATI once again is cheating using the drivers?

or probelly the driver brings the Picture quality down.. and performance high...

which for nvidia... it doesn't change!

[jbirney]

I don't get it....

There is no cheating by ATI with these drivers. It fixes a bug with vertex data management.