I know that HDR+AA is possible on ATI implemention, and thats a pro but why else. I currently own a 6800GT and I don't want to start a flaming war, I just want to understand. I am looking to buy my first high-end card so this is why I started this thread. Is ATI's PS 3.0 better because the X1800XT is clocked so fast or just because they work closely to Microsoft.
Can someone tell me why ATI's PS 3.0 is better than Nvidia's?
- Thread starter Redeemer
- Start date
From my understanding it's because they implement it differently.
Redeemer said:
Totally different execution style.
nVidia seems to have treated PS3.0 in their first iteration as more of a PS 2.0 with some branching added on. (For the more technically inclined I know this is a simplification.) This shows particularly with the branching stalling the rendering pipeline for long periods such that unless skipping a large chunk of shader code then the branch ends up being a performance loss.
ATi have gone for a more decoupled approach that allows them to hide the latencies of branching more effectively. It isn't that the shader doesn't stall, but more that the GPU has the ability to set aside that particular shader while the data it is waiting for is fetched. The GPU then gets on executing something else rather than stalling the pipeline completely.
The impression I get is that a good healthy mix of branching, arthimetic calculation of pixel values as well as straight texturing is what the ATi part is aimed squarely at - favouring longer and complicated shaders. nVidia's current parts seem to favour shorter less convoluted shaders and an emphasis on straight texturing performance.
Which is better? nVidia's approach is very suitable for right now but I think I give the edge to ATi for titles coming in a years time or so.
Redeemer said:
I'm not sure people are concluding anything, but future titles may favor the X1800 because of advantages in dynamic branching. There's also the case where the X1800 does outperform the g70 in fear by a pretty big margin, but I wouldn't put a lot of emphasis on fear numbers until its a released product.
http://www.tgdaily.com/2005/10/07/stanford_gpgpu/index.html
This tidbit from toms might shed some light...
This tidbit from toms might shed some light...
Bouncing Zabaglione Bros.
Legend
Unreal Engine 3 games arn't far off, and they will be everywhere. Then there are upcoming games like Stalker, Alan Wake, etc. 2006 will be the year for SM 3. I think dynamic branching may give a significant speed/quality boost to these titles. By that I mean these titles will be very graphics heavy, and without SM3 to keep the speed up, you're going to see a drop in visual quality to keep the framerate playable.rwolf said:
Bearing in mind the success of NV4X/G70 is it likely we'll even see much dynamic branching used in games currently undergoing development? The developers obviously will know that dynamic branching cripples the performance on NV's current architectures so I'd expect them to avoid using this as much as possible.
The take-up of SM2.0 was probably slower due to all the GF3/4 cards on the market and also the poor performance of NV3X series so I wonder if a similar thing might happen with dynamic branching?
I'd imagine that ATI's lack of SM3.0 support until now has also delayed the usage of this shader model to a certain extent too, however!
The take-up of SM2.0 was probably slower due to all the GF3/4 cards on the market and also the poor performance of NV3X series so I wonder if a similar thing might happen with dynamic branching?
I'd imagine that ATI's lack of SM3.0 support until now has also delayed the usage of this shader model to a certain extent too, however!
Bouncing Zabaglione Bros.
Legend
Mariner said:
Developers have been saying since NV40 that SM3 makes their jobs easier and simpler, and if that's true, I expect SM3 to take off for that reason alone. Developers like to get good results for mimimal work.
I'm sure that everyone is expecting Nvidia to sort out their branching performance hit by next year when they bring out the .09 version of G70.
Bouncing Zabaglione Bros. said:
Are they? I'm not. I wouldn't be shocked or anything, but I consider it on balance to be less likely rather than more likely --at least if by that you mean branching performance equal to ATI's. At least not if they are also upping PS to 32, as is the rumor. Why do I say that? 1). Because it appears they already increased it somewhat for G70, and I suspect that was the end of the "easy wins" on that front for them. 2) Because ATI doubled transistors without increasing functional units to create "SM 3.0 done right".
Maybe they are mondo wizards at NV, but I'm not expecting them to figure out how to get north of 400m transistors on 90nm on their first high-end part, nor how to get a signficant performance increase in this area without spending signficant transistors on it.
Bouncing Zabaglione Bros.
Legend
geo said:
There's no doubt that going to .09 with such a complex chip will be difficult for Nvidia. There's no doubt that ATI has upped the stakes dramatically as far as SM3.0 branching performance goes (one of the main reasons for SM3.0 at all).
I just don't know what Nvidia would be thinking if after having SM3.0 in NV40 and then G70 and promoting it so heavily, they don't fix one of the most important parts of SM3.0 that is severely lacking on their chips with their third attempt.
I doubt there is much Nvidia can do now to change their design for a .09 G70, so if they were expecting to get a free pass from R520 and planned accordingly, they will be in big trouble six months from now. Nvidia will get slaughtered as soon as we see games and benchmarks that use the SM3.0 they've been plugging so heavily.
Similar threads
