Anand's HL2 benchmarks

[Natoma]

http://www.anandtech.com/video/showdoc.html?i=1863&p=10

Here we see something very interesting, and something we haven't really seen before - the Radeon 9600 all the way up to the Radeon 9800 Pro all perform within 12% of each other. This is because with shader-heavy games such as Half-Life 2, the bottleneck is no longer memory bandwidth - rather it is pure computational power; basically how quickly these GPUs can process through those shader programs.

I would think that it's not the shader computations that make the 9600 Pro perform within 12% of the 9800 Pro, but CPU limitation. Considering the 9600 Pro has half the shader units of the 9800 Pro, if this were really a shader limited scenario the 9800 Pro should be screaming away from the 5900 Ultra and the 9600 Pro should it not?

Damn... CPU limited with a 3.0C.. Dear Lord..... Oh well. "Free" AA/AF everyone!!

As long as you're using a 9500+ card though. yay centroid sampling.

You know though, those faulty D3 benchmarks from anand were up for a good 3-4 weeks. I wonder how long these numbers stay at the top of the list?

 

[Mintmaster]

Wasn't a bad read but it would have been MUCH better with some screenshots to show the difference between the various dx modes. Particularly the difference between dx9 and mixed mode. Hint. Hint. Dave? Anything coming up at beyond3d?

My guess is that there are NDA issues. Valve is probably not quite ready to have Anand putting up arbitrary screenshots.

 

[Natoma]

Lol Natoma....what are you talking about?

The graphs were all the same for about 20 minutes. I noticed a few comments in the forums link for the article and other people were seeing the same thing. Don't mind me.

 

[jjayb]

I saw the same thing Natoma. Had me really confused at first till I realized they had the same graph on every page.

 

[Mintmaster]

I would think that it's not the shader computations that make the 9600 Pro perform within 12% of the 9800 Pro, but CPU limitation.

You're totally right. The best way to see CPU/geometry limitations in a GPU review is look at resolution changes. The 9600+ all have similar framerates at both 1280x1024 and 1024x768. The 5900, though, does not

I wouldn't worry about the CPU usage though. That can be corrected by Valve by reducing physics iterations and other optimizations.

 

[Natoma]

I would think that it's not the shader computations that make the 9600 Pro perform within 12% of the 9800 Pro, but CPU limitation.

You're totally right. The best way to see CPU/geometry limitations in a GPU review is look at resolution changes. The 9600+ all have similar framerates at 1280x1024. The 5900, though, does not

I wouldn't worry about the CPU usage though. That can be corrected by Valve by reducing physics iterations and other optimizations.

You know something? I want to see the Athlon64 running this game. Look here at Gunmetal DX9 benches with a 9800 Pro 128MB and 3.6 Cats.

http://www.anandtech.com/cpu/showdoc.html?i=1856&p=5

Benchmark 1
P4 3.0C - 36.91fps
Athlon64 2.0Ghz - 51.61fps

Benchmark 2
P4 3.0C - 27.61fps
Athlon64 2.0Ghz - 42.62fps

Holy Jeebus CPU limitations. Sigh. And here I thought my system would be worth something for HL2 a couple of months ago with this 3.0C CPU and 9800 Pro 256MB.

/me bashes the CPU in and hopes for Prescott to perform much better...

 

[Dave H]

Here we see something very interesting, and something we haven't really seen before - the Radeon 9600 all the way up to the Radeon 9800 Pro all perform within 12% of each other. This is because with shader-heavy games such as Half-Life 2, the bottleneck is no longer memory bandwidth - rather it is pure computational power; basically how quickly these GPUs can process through those shader programs.

Honestly. There ought to be some certification test you need to pass before you're allowed to review hardware online. The sheer ignorance on display here is breathtaking--but perhaps not so much as the number of times I've already seen Anand's brilliant "memory bandwidth is no longer a bottleneck" insight repeated online.

In any case, the 9600 results on all the benchmarks are extremely suspicious, which is to say definitely incorrect. The first two show no difference whatsoever between 1024*768 and 1280*1024. And no, it's not because the 9600 scores are CPU limited, as the 9800 and 9700 are doing significantly better. It's a bug, either on the part of HL2/ATI drivers or on the part of Anand. I'll let you guess which one I suspect...

 

[Natoma]

And no, it's not because the 9600 scores are CPU limited, as the 9800 and 9700 are doing significantly better. It's a bug, either on the part of HL2/ATI drivers or on the part of Anand. I'll let you guess which one I suspect...

He probably forgot to change the resolution for the 9600 Pro. Either that or someone screwed up the spreadsheet.

 

[horvendile]

In any case, the 9600 results on all the benchmarks are extremely suspicious, which is to say definitely incorrect.

I noticed that too!
Could one possible explanation be that it is geometry limited, while the 9700+ are shader limited?
It should be possible to make some rough calculations to see if that's remotely likely. (Er... maybe.) I'll do it if I can justify continue doing non-work things. Unless someone convinces me that it's futile.

Is it reasonable to assume that all R(V)3X0 have the same geometry and pixel shading power per pipe and clock?

 

[NapalmV]

AA/AF,

http://www.gamersdepot.com/hardware/video_cards/ati_vs_nvidia/dx9_desktop/hl2_followup/003.htm

Now that's funny !

 

[gokickrocks]

AA/AF,

http://www.gamersdepot.com/hardware/video_cards/ati_vs_nvidia/dx9_desktop/hl2_followup/003.htm

Now that's funny !

the first page shows the IQ differences

http://www.gamersdepot.com/hardware/video_cards/ati_vs_nvidia/dx9_desktop/hl2_followup/shotdx.htm

 

[Trawler]

:? Gabe seems very passionate about this. I'm really starting to wonder if he's going to up the ante.

Depending on whether nVidia play ball or not and remove their naughty optimizations in the final release of Det50, it could well be that Gabe decides to pump out a little driver detection routine through Steam.

What this driver detection routine could do is up to how pissed Gabe is with NV, but an alert telling the user that the drivers they're using will degrade IQ would be more than enough to push nv onto their back foot (not like they aren't already).

This isn't FutureMark, this is the creators of Half Life 2. NVidia has chosen the wrong guy to go up against, and he's going to make 'em pay for it.

All IMHO of course.

 

[Bjorn]

the first page shows the IQ differences

http://www.gamersdepot.com/hardware/video_cards/ati_vs_nvidia/dx9_desktop/hl2_followup/shotdx.htm

Those were rather crappy (and boring) screenshots imo.

 

[Bouncing Zabaglione Bros.]

:? Gabe seems very passionate about this. I'm really starting to wonder if he's going to up the ante.

Depending on whether nVidia play ball or not and remove their naughty optimizations in the final release of Det50, it could well be that Gabe decides to pump out a little driver detection routine through Steam.

It's already been said by Valve that it will be very difficult to cheat in online games of HL2 because with Steam they can issues fixes very quickly - several times an hour if they want to. How easy do you think it will be for Valve to stymie any Nvidia shader detection by changing a few bytes here or there to stop Nvidia's drivers from detecting the code? Very easily - everytime someone logs onto Steam, a few extra bytes here and there, and Nvidia's cheats are disabled.

 

[horvendile]

In fact, the discrepancy between the results of gamersdepot compared to those Anand obtained rhyme fairly well with the difference in CPU speed, perhaps indicating a CPU limitation for the 9800 Pro.

But wait a minute... if it was a CPU limitation then there wouldn't be any noticeable decrease in fps as we up the resolution, but there is. Of course, the shader / fillrate limitation could also be there somewhere, but I feel that I'm in deep water now. (Non-Fresnel refracted water, since I don't have a DX9 card.)

 

[Bjorn]

But wait a minute... if it was a CPU limitation then there wouldn't be any noticeable decrease in fps as we up the resolution, but there is. Of course, the shader / fillrate limitation could also be there somewhere, but I feel that I'm in deep water now. (Non-Fresnel refracted water, since I don't have a DX9 card.)

I'm guessing that some parts of the map/demo are CPU limited and some are more shader limited. Thus, raising the res would lead to lower framerates in the shader limited areas and of course lower overall framerates.

As an example, one of the demos at Anands had the 9600 Pro performing even a bit faster then the 9700 Pro and another one showed it to be closer to half the speed. The one were they scored almost the same was probably almost entirely CPU limited. Or, hmm, maybe vertex shader limited ?

 

[StealthHawk]

the first page shows the IQ differences

http://www.gamersdepot.com/hardware/video_cards/ati_vs_nvidia/dx9_desktop/hl2_followup/shotdx.htm

Those were rather crappy (and boring) screenshots imo.

It was said that High Dynamic Range effects were not yet implemented in the maps used for the HL2 benchmark.

 

[Bjorn]

It was said that High Dynamic Range effects were not yet implemented in the maps used for the HL2 benchmark.

I also watched the latest video. And i don't know, to me it looks like: take the UT2003 engine, slap on some shaders for better looking surfaces. Add some polygons. And voila, HL2. Maybe that's because my main problem with the current 3D games/engines is the crappy lighting/shadowing and HL2 doesn't seem to improve that much in that area. From what i have seen yet at least. I might also be a bit spoiled since i tried the leaked Doom3 alpha.

 

[horvendile]

And i don't know, to me it looks like: take the UT2003 engine, slap on some shaders for better looking surfaces.

Maybe, but they use the Fresnel equations for water surfaces... that alone may sell me a DX9 card!

 

[WaltC]

I also watched the latest video. And i don't know, to me it looks like: take the UT2003 engine, slap on some shaders for better looking surfaces. Add some polygons. And voila, HL2. Maybe that's because my main problem with the current 3D games/engines is the crappy lighting/shadowing and HL2 doesn't seem to improve that much in that area. From what i have seen yet at least. I might also be a bit spoiled since i tried the leaked Doom3 alpha.

I was very impressed with the movie--to me, the textures and reflectivity of the scenes look more lifelike than anything else I've seen. The monster on the roof looked real, I thought. It was also obvious that FSAA was used this time, as well, which made things look a lot better than the previous movies. Plus--it's a movie--not a demo, so a little is lost in the translation.

Speaking of lighting, which scenes in the D3 alpha are outside on a rooftop in broad daylight with the sun peering down through the clouds?.... I think there will be plenty of places in this game that are dark and shadowy like D3. Right now, HL2 is *the* game I want. Waiting for D3 until next year doesn't bother me at all...and I wouldn't be surprised if upon seeing HL2 ID doesn't do a bit more work on the game prior to release (which is why I think D3 was put back in the first place--just MO...)