UT 2003 GPU Shootout Part II at Anandtech

[Basic]

pascal:

The first error is that you asume that a 1.6GHz processor is twice as fast as a 800MHz processor, when everything else (including FSB) is kept constant.

And then you have misused Amdahl's law.
The formula is right, the way to use it isn't.

What you're missing is that a lot of the GPU calculations are done in parallel with CPU, so you can't get anything out of Amdahl's law regarding how it will behave in a different resolution.

 

[pascal]

Hi Basic

I know that a 1.6GHz system is not twice as fast as a 800Mhz system. It was a simplification/aproximation to show a point.

Yes, there are a lot of parallel action but see the GF4 curve again, it is the familiar curve before saturation. They (CPU and GPU) behave like two serial tasks. Try to plot the curve using Amdhal´s law.

 

[pascal]

Basic

One more proof. See this two benchmarks:http://www.anandtech.com/video/showdoc.html?i=1647&p=12

http://www.anandtech.com/video/showdoc.html?i=1647&p=13

From the scalling analysis with 800x600 high detail (+-145fps) GF4Ti4200/Athlon 800Mhz: CPU 50% and GPU 50%

Lets predict the fps with 1024x768 high detail GF4Ti4200/Athlon1.73GHz.
I will use some shortcuts/normalizations with Amdhal´s law.

CPU: 50% / (1.73/.8 ) = 23%

GPU: 50% / ((800x600)/(1024x768)) = 81% edited minor order correction

Total for 1024x768 high detail GF4Ti4200/Athlon1.73GHz = 104%
Total for 800x600 high detail GF4Ti4200/Athlon1.73GHz. = 73%

Framerate at 800x600 high detail GF4Ti4200/Athlon1.73GHz (measured) = 196fps

Framerate at 1024x768 high detail GF4Ti4200/Athlon1.73GHz (measured) = 133fps

Using the last as the startpoint then 133.33x (104%)/(73%) = 190 fps

Sumarizing, 190 fps predicted and 196fps measured (prediction with error lower than 3%) 8)

 

[Nagorak]

People claiming there is some "Nvidia bias" in UT2003 really get on my nerves. Look at the benchmarks from the first comparison, you can see the R8500 hangs with the GF3 just fine in intensive situations. If it's slower when nothing is happening, who cares?

This is just ridiculous. If you're staring at a blank wall and you get 200 fps with a GF3 compared to 150 fps with the Radeon 8500, what difference does it make? If both cards perform the same under action, that's all that matters.

As to this comparison...I have to wonder WHY they tested in 800*600? If the point of the comparison was to see how much the graphics card impacted performance vs how much the CPU impacted performance, why did they run in a low resolution that nobody uses?

Basically this comparison was totally useless to me. It more or less showed that the graphics card was the bottleneck in almost all situations at 800*600. So I guess I can assume that the graphics card is even more the bottleneck at higher resolutions. But I don't know for sure... basically this performance test seems pretty pointless.

Why they didn't use 1024 resolution is beyond me, does anyone here actually run in 800*600???

 

[pascal]

...
Why they didn't use 1024 resolution is beyond me, does anyone here actually run in 800*600???

For your first question, well with 1024x768 probably all cards (including the GF4Tixxxx) will be GPU limited and all scalling curves will be almost flat.

At 800x600 the GF4Tixxxx diferentiate themselves from the rest (nice marketing). edited: probably it was coincidence

For your second question sometimes I play at 800x600. The secret is my small 15" trinitron monitor with a physical grid of 1024x768 and .25 dot pitch edited: I use my WinXP at 1024x768.

edited: finally they should have done this CPU scalling with two resolutions 640x480 (for predictions purpose) and 1024x768 (very popular resolution). But anyway I liked their hardwork.

 

[Randell]

A lot of people will be playing UT2003 at 800x600 anyway, if they wish to sustain fps above 40-50 I would guess.

 

[Seiko]

Most ppl will be CPU bound!

A lot of people will be playing UT2003 at 800x600 anyway, if they wish to sustain fps above 40-50 I would guess.

I'm afraid unless Sweeny has really shifted the emphasis back to the GPU and not the CPU a lot of people will be CPU bound. I have an old CeleronII 850 and 8500. Unless I utilise FSAA I'm totally CPU bound and can switch from max to min settings any resolution with no effect! The only problem is as soon as bots enter the arena my framerates drop to the low 20s/high teens! So I guess I'll be needing that CPU upgrade just to play!

Without starting a UT/Quake3 debate I do admire JC for how he managed to keep Q3 GPU bound. I plugged in the 8500 over an V5500 and my framerates doubled. I can't say the same for UT nor UT2003.

Ah well, I guess if your AMD based anyway then it's probably a cheaper upgrade if you're CPU bound and not GPU bound?

 

[Teasy]

But you should see it on Kyro. Looks and runs terribly

Are you talking about Kyro 1 or II?

 

[Bambers]

Re: Most ppl will be CPU bound!

I'm afraid unless Sweeny has really shifted the emphasis back to the GPU and not the CPU a lot of people will be CPU bound. I have an old CeleronII 850 and 8500. Unless I utilise FSAA I'm totally CPU bound and can switch from max to min settings any resolution with no effect! The only problem is as soon as bots enter the arena my framerates drop to the low 20s/high teens! So I guess I'll be needing that CPU upgrade just to play!

Without starting a UT/Quake3 debate I do admire JC for how he managed to keep Q3 GPU bound. I plugged in the 8500 over an V5500 and my framerates doubled. I can't say the same for UT nor UT2003.

Ah well, I guess if your AMD based anyway then it's probably a cheaper upgrade if you're CPU bound and not GPU bound?

I think the game is alot faster since that leak. I was testing yesterday and comparing against anands results.

With medium details at 800x600 I get ~50fps with no bots and 30fps with 3 bots on asbestos. Anand gets ~140fps on GF4 (even MX) cards on a 800Mhz XP (nearest to my 1Ghz tbird). Thats a very big improvement.

 

[Dave Glue]

Re: Most ppl will be CPU bound!

I'm afraid unless Sweeny has really shifted the emphasis back to the GPU and not the CPU a lot of people will be CPU bound. I have an old CeleronII 850 and 8500. Unless I utilise FSAA I'm totally CPU bound and can switch from max to min settings any resolution with no effect! The only problem is as soon as bots enter the arena my framerates drop to the low 20s/high teens! So I guess I'll be needing that CPU upgrade just to play!

With the prices of CPU's these days, does it make much of a difference as opposed to ugprading your graphics card? An XP1800+motherboard combo can be had for ~$200 US, things are somewhat different as opposed to when UT hit the shelves. A Celeron 850 is quite long in the tooth for a lot of games, did you really expect a game that pushes far more polygons than UT to not have a large demand on the CPU?

 

[pascal]

Most slowdowns are because bots?

Are the UT2003 Bots better/smarter than Q3TA bots? Why the slowdown?
Is it a buggy?

 

[Bjorn]

Most slowdowns are because bots?

Are the UT2003 Bots better/smarter than Q3TA bots? Why the slowdown?
Is it a buggy?

Maybe not optimized yet ? (bots)

Or maybe both buggy and not optimized.

 

[jb]

I think the game is alot faster since that leak. I was testing yesterday and comparing against anands results.

Yeap. That was one of the first thing out of Mark Rein during the UT mod summit. He was pissed as hell about the demo leak. Mainly do the the fact it was before ANY optimizations had been done and it was a very incomplete build (lacking the karma physics for one of many big things it was missing). Some poeple have already passed judgement on the game due to this leaky and thats sad as it was such crap to what we got to play with.

Are the UT2003 Bots better/smarter than Q3TA bots? Why the slowdown?

Pascal,

Steve P. had a chance to sit down with the coders and talk about the AI changes from UT (as he did most of that work). I am not sure you knew this but the getting the bots to do anything other than play DM was a night mare. It took me weeks to teach them how to play King of the hill style of games. Often the topic of bots comes up on the UT coding forums. Very tough. Steve said that will change. For UT2003 the bots have three levels of AI with 2 of the 3 that are at the coders disposable in case he/she wants to do anything different. There is a low level bot AI that handles the "basic" stuff like moving, targeting enemies and firing. Then there is what he termed as a team AI. This was tied into what he called the level AI so I will try to explain both. The level AI controls the objectives that the bots are trying to do (be it defend the flag, hold the dom point, or find the next hill). This also will take into account the objectives of the game and actually change the way it plays the game depend on how the game is going. The team AI works to help the bots work together in a group. For example the Game AI may tell the bots to guard the flag as well as get the enemies flag. The team AI says OK I will have 2 stay hear and guard our flag. And then I will send the other 2 off to get their flag. Steve said that he has seen the AI during a CTF game, when it was winning the game pull all of its bots back to play Defense during the last minute! That's cool. Also keep in mind for a DM game, the Team AI knows its a team of only 1. Again this is more or less what we got from Steve. I have yet to try it out for my self but should soon. I am looking forward to get my hands on this after banging my head for so long trying to get the UT bots to do stuff.

Jb

 

[Ozymandis]

Actual the 927 leak does have Karma physics. I think. Although I don't think it's totally in there yet.

But you should see it on Kyro. Looks and runs terribly

Are you talking about Kyro 1 or II?

I had to drop in my old Evil Kyro 64meg card, because the GeForce3 wouldn't work with the latest official drivers. Got it working with the GF3 after installing the latest leaked drivers, but anyways:

The Kyro has artifacting. Serious artifacting. In medium and high texture detail there were the largest problems. Textures wouldn't load in high-detail scenes when you were moving, it just showed blank polygons. The HUD didn't work unless you stared down at the floor. Water was invisible on the Kyro as well. And of course framerates were totally unplayable.

 

[Teasy]

Ah I thought so, I'd expect pretty bad framerates with a Kyro 1 board, Kyro II is about 55% faster then Kyro 1. Although it also depends what sort of CPU you have, I suppose you need a decent CPU to get good performance. With a fast CPU you may get a half decent framerate with a Kyro 1 at lower res (640x480, and even maybe 800x600 in low detail).

As for your problems with artifacts ect, obviously Anand isn't having those problems with the Kyro so perhaps your using old drivers? Although you do have the worst made Kyro board around (Evil Kyro from Powercolor) so maybe its that. Often the Powercolor Evil Kyro would get bad artifacts in some games at defualt clock speed, dropping the GPU/ram speed to 160mhz (or perhaps it was 165, I can't remember) was reported to fix that problem.

 

[Hellbinder]

In Army ops which is very detailed and uses the new Engine my new system averages between 45-130 FPS at 1280x1024... with "everything enabled" ... what little of it you can control that is...

The only time i see any serious slowdown is when a buch of smoke grenaged are tossed... in the thick of it I have hit about 35 FPS..

Yes, i do know how to enable the Frame counter in that game its

Stat FPS

 

[Ozymandis]

Ah I thought so, I'd expect pretty bad framerates with a Kyro 1 board, Kyro II is about 55% faster then Kyro 1. Although it also depends what sort of CPU you have, I suppose you need a decent CPU to get good performance. With a fast CPU you may get a half decent framerate with a Kyro 1 at lower res (640x480, and even maybe 800x600 in low detail).

As for your problems with artifacts ect, obviously Anand isn't having those problems with the Kyro so perhaps your using old drivers? Although you do have the worst made Kyro board around (Evil Kyro from Powercolor) so maybe its that. Often the Powercolor Evil Kyro would get bad artifacts in some games at defualt clock speed, dropping the GPU/ram speed to 160mhz (or perhaps it was 165, I can't remember) was reported to fix that problem.

I'm using the newest drivers, on a Tbird 1.3 ghz. *shrug* I don't know, maybe it's just that the compatibility in this leaked build is poor. I don't have problems in any other games with the Kyro.

We'll see how it turns out. Personally, I could care less. My main card is a GeForce3 at more than Ti500 speeds. That should be fine for UT2k3 8)

 

[Seiko]

I think the game is alot faster since that leak. I was testing yesterday and comparing against anands results.

Yeap. That was one of the first thing out of Mark Rein during the UT mod summit. He was pissed as hell about the demo leak. Mainly do the the fact it was before ANY optimizations had been done and it was a very incomplete build (lacking the karma physics for one of many big things it was missing). Some poeple have already passed judgement on the game due to this leaky and thats sad as it was such crap to what we got to play with.

Are the UT2003 Bots better/smarter than Q3TA bots? Why the slowdown?

Pascal,

Steve P. had a chance to sit down with the coders and talk about the AI changes from UT (as he did most of that work). I am not sure you knew this but the getting the bots to do anything other than play DM was a night mare. It took me weeks to teach them how to play King of the hill style of games. Often the topic of bots comes up on the UT coding forums. Very tough. Steve said that will change. For UT2003 the bots have three levels of AI with 2 of the 3 that are at the coders disposable in case he/she wants to do anything different. There is a low level bot AI that handles the "basic" stuff like moving, targeting enemies and firing. Then there is what he termed as a team AI. This was tied into what he called the level AI so I will try to explain both. The level AI controls the objectives that the bots are trying to do (be it defend the flag, hold the dom point, or find the next hill). This also will take into account the objectives of the game and actually change the way it plays the game depend on how the game is going. The team AI works to help the bots work together in a group. For example the Game AI may tell the bots to guard the flag as well as get the enemies flag. The team AI says OK I will have 2 stay hear and guard our flag. And then I will send the other 2 off to get their flag. Steve said that he has seen the AI during a CTF game, when it was winning the game pull all of its bots back to play Defense during the last minute! That's cool. Also keep in mind for a DM game, the Team AI knows its a team of only 1. Again this is more or less what we got from Steve. I have yet to try it out for my self but should soon. I am looking forward to get my hands on this after banging my head for so long trying to get the UT bots to do stuff.

Jb

I don't think people are jumping to conclusions. Certainly ppl are trying to estimate a performance level using a very early and buggy version. You could argue that it invalidates all results and the final version will be umpteen times better and quicker for all. Personally I don't believe that will be the case. Sure. many code optimisations will be finalised but as the demo already lacks karma physics alot of the optimisations will simply ensure that the same level of performance is maintained even running the karma modules.

Of course all this is speculative and we'll just have to wait and see what transpires when the official demo is released. I still feel alot of ppl with sub Athlon/P4s will be needing the upgrade to play any level with more than one bot.

Of course I hope I'm wrong so I can save my pennies.

 

[Seiko]

Re: Most ppl will be CPU bound!

I'm afraid unless Sweeny has really shifted the emphasis back to the GPU and not the CPU a lot of people will be CPU bound. I have an old CeleronII 850 and 8500. Unless I utilise FSAA I'm totally CPU bound and can switch from max to min settings any resolution with no effect! The only problem is as soon as bots enter the arena my framerates drop to the low 20s/high teens! So I guess I'll be needing that CPU upgrade just to play!

With the prices of CPU's these days, does it make much of a difference as opposed to ugprading your graphics card? An XP1800+motherboard combo can be had for ~$200 US, things are somewhat different as opposed to when UT hit the shelves. A Celeron 850 is quite long in the tooth for a lot of games, did you really expect a game that pushes far more polygons than UT to not have a large demand on the CPU?

Nope in fact I believe it raises a very valid question for game developers. Whats the best % split between CPU and GPU? In fact perhaps this should be started in another thread?

 

[Bambers]

You could argue that it invalidates all results and the final version will be umpteen times better and quicker for all. Personally I don't believe that will be the case. Sure. many code optimisations will be finalised but as the demo already lacks karma physics alot of the optimisations will simply ensure that the same level of performance is maintained even running the karma modules.

I think performance will increase a large amount.

I get around 30fps on the 927 build. Anand on a (presumably) newer build gets at least 140fps for an equivilent system.