ATI Hierarchical-Z issue with Doom 3

Poor subject ...

It might be ineffective, when you said it had an issue I was assuming you meant it created artifacts though.

 

Personally, I'd like to know where they are basing a lot of the infomration from - there doesn't appear to be any testing involved, and from the translated text I have a few questions over some of the thing said (that may be the translation though).

 

Interesting article. I wonder though how that will affect radeon 8500/9000, could this game actually be faster on a 9000 than on a 8500? Everybody just assumed that the 8500 is faster mostly because it has 2 texture units, but personally I always thought that maybe it's in fact Hierarchical-Z which makes more of a difference in practice. But if Hierarchical-Z doesn't quite work in DoomIII anyway, that should be an interesting comparison...

 

If I had to make a wild guess (and it would pretty much be pulled out of my Ass) at the issue on ATI cards, I'd say it is most likely an issue with early Z reject and the way that Doom 3 renders.

And further speculating it is most probably related to the swapping of Ztest modes in the rendering path. I seem to remember an ATI PDF that implied that this was bad with Hyper Z. AFAIK in Nvidias early Z implementation there is no affect from this.

I base this purely on the scale of the disparity and what little I know of ATI's architecture and the rendering methods that Doom3 uses.


It could just as easilly be a an OpenGL driver with a lot of overhead, and Doom3 making a LOT of draw calls. But I doubt this particular scenario.

 

hasn't this been discussed here before? I thought this was fixed or improved with r350.

 

In Gabe's words : "God willing..."

 

I tested my 8500 64MB (275/300) and my 9000 Pro 128MB (275/250) at 640x480 & low quality. The 9000 certainly did run the game faster, but I chalked that up to the onboard memory size differences.

 

Radeon 8500 also has dual vertex shaders. 9000 has only 1. So, the 8500 should theorectically have more texel fillrate and more polygon pushing power.

I remember Carmack mentioning that certain IHVs had mispredicted the need for granularity in the memory controller. If I recall right, the 8500 reads in rather large chunks (256bit?) compared to other cards (GF4's crossbar). I think this was causing performance problems in Doom3. ATI was planning for lots of big textures, but that's not the big drain in Doom3.

Also, a 128MB 8500 has memory interleaving that 64MB cards do not. This gives a decent boost in performance. I believe 128MB 9100s and perhaps 9000s do not enjoy this advantage.

 

and for my mobility 9200 this means?
Besides that it will run like ace.

 

Does the Doom3 engine do something that's inherently bad for ATI's architecture? Does it look like this engine was designed very much around a NV chip (or with NV's guidelines/ideas so they could design a chip around it)?

I know John usually gets together with the IHVs and discusses how future chips can be better. Maybe NV sat down with him a lot more than ATI and gave him lots of free hardware.....

 

What a nice conspiracy theory. Maybe ATI hardware just doesn't have as much stencil fillrate. Ever thought of that?

Dave,
the article points towards ATI's hierarchical Z and its problems with reversal of the depth test mid-scene. Sounds feasible.

aths attended certain NDA briefings and is clever enough to ask some of the right questions. I'd trust him on this one.

 

But can we trust him to get us a good English translation of his latest investigation?

 

Leonidas (site owner) promised that a translation was in the works. Dunno when. I doubt aths himself will do it. Probably nggalai. If that's the case it's bound to become a good translation

 

Ah stop the NV/ATi crap please. Runs extremely nice at High Settings at 1024x768 on R9600 Pro (pc'ed at 500/700). You just have to tweak cache size and some other stuff in D3 config.

 

Doom 3 and my ATI 9800 Pro with 128 megs of ram

Doom 3 runs like crap on my ATI Radeon 9800 Pro
with 128 Megabytes of ram.

Pentium 4 2 Ghz
512 Megs of DDR Ram
SB Audigy 2 audio card (Latest drivers)
Windows XP (all latest updates minus DX 9.0c)
Direct X 9.0b
ATI Radeon 9800 Pro with 128 megs ram (with one
CRT monitor setup and the latest drivers available)
3 mbps download / 384kbps upload for DSL
100 mbps Ethernet card

I have my settings for Doom 3
at 1024x768x32 with Medium Details and
not even playing multiplayer and its very slow.

If more than one demon is on the screen at the same
time its like a slide show, thats how slow it is.

oh and I have had Vsync off or on in the game
(setting in game not in videocard drivers and it
doesn't make any difference as far as speed, its
still super slow).

 

I am NOT trying to start some bullshit anti-NV campaign here. I am far from biased towards one vendor. I will buy whatever performs. ATI has been leading the pack for a couple years, and now Doom3 comes up and we see this fairly significant difference between 6800 and X800.

I didn't know there was a difference in stencil fillrate......How can you determine how much stencil fillrate a chip has?

 

Proforma, try this to improve performance.

 

You know, RV3x0 lacks Heirarchical Z. I guess dropping that from the core for size considerations was a good call in some backward way after all.

My laptop's 9600 64mb is the little VPU that could...445MHz of mosquito buzzing energy in that little chip. Too bad the RAM so so pathetically slow at 230MHz....wish that GDDR2-M had arrived.

It is something though how much better my 9700PRO 128MB @ 385/350 cuts thru Doom3. I love R300. How many graphics chips can blast through a state of the art game 3 years after release? My R100 and G400 lasted like a little over a year each.

 

Just saw this at Firingsquad:

http://www.firingsquad.com/hardware/doom3_perf/page11.asp

Doesn't say anything about why it isn't used though, if it's true that is.