Opinions needed on this Interview
- Thread starter JoshMST
- Start date
I am not sure if ATI have some kind of dynamic move for the splitline. I have not see anything but the driver is very large and it is possible that i lost the trace.
If I have understand the nVidia solution right they assign rectangles to each card. This make a vertical splitline possible. IMHO they use a horizontal line because it make the scan out over the bridge easier.
If I have understand the nVidia solution right they assign rectangles to each card. This make a vertical splitline possible. IMHO they use a horizontal line because it make the scan out over the bridge easier.
Well, nVidia's dynamic load-balancing would almost have to be reactive, not proactive, no? Seems to me it'd only react after the current split stumbled for a (few) frame(s), so it probably wouldn't be as effective in boosting the minimum framerate as it would the average. It also seems (in this Extreme Layman's opinion) that a vertical split would sidestep most of the issues that a horizontal split would encounter. In the games where a variable or asymmetric split would be preferable, it would be due to the *hor*izon, not the *ver*izon.trinibwoy said:
It's probabaly more likely you're seeing sky above and ground below than ground on the left side and sky on the right (say, near a mountain?).I don't think so. A single fixed vertical split (or any other direction) is by far not enough to even the workloads out. It happens far too often that part of the screen is blocked by a wall, or an explosion/smoke/complex object is only visible in one area of the screen.Pete said:
You really need to partition the screen into loads of small segments and hope for the workload to average out, or do a dynamic splitting.
Dynamic splitting works reasonably well because there can't be too much changes frame-to-frame since otherwise the motion would become jerky.
Pete said:
Well there was some discussion a while back on how this might work. It is reactive as you say but it may be recalculated per-frame. I think this would improve minimum framerates as well since the prior frame's load-balancing ratio should provide a very good approximation for the next frame's in most situations (assuming a relatively high-framerate in the first place).
Pete said:It also seems (in this Extreme Layman's opinion) that a vertical split would sidestep most of the issues that a horizontal split would encounter. In the games where a variable or asymmetric split would be preferable, it would be due to the *hor*izon, not the *ver*izon.
I agree with Xmas above. In indoor areas there is a high probability that one side of the screen has a lot less geometric detail. Of course a horizontal split also has issues in outdoor areas where the sky is visible as you pointed out.
Does every frame need to be finished?
Let's say, that one card can't finish all it's tiles, would it be able to drop the output for one frame to work on the next one because simply, it has no time to finish that frame?
It would then signal the master card that workload is too high and it wants less work for the next frame.
I doubt if there would be any "jerking" since you'd miss out on two frames (max I guess.) where a number of tiles would be missing.
Especially if you consider you'd get higher minimum (and average) framerates, missing, say 20 tiles in a 40fps environment would hardly be noticeable.
Dumping tiles in a high fps environment to possitively affect the minimum framerate and load balancing seems like a worthwile tradeoff.
Let's say, that one card can't finish all it's tiles, would it be able to drop the output for one frame to work on the next one because simply, it has no time to finish that frame?
It would then signal the master card that workload is too high and it wants less work for the next frame.
I doubt if there would be any "jerking" since you'd miss out on two frames (max I guess.) where a number of tiles would be missing.
Especially if you consider you'd get higher minimum (and average) framerates, missing, say 20 tiles in a 40fps environment would hardly be noticeable.
Dumping tiles in a high fps environment to possitively affect the minimum framerate and load balancing seems like a worthwile tradeoff.
Jawed said:
Well, as always without any warranty.
It doesn't look like. There is one part for vertical split and one for horizontal splitline calculation but if i understand it right the can not used at the same time. Using two splitline at the same time with only two chips make the imagecombinding more complicated. I am talk about 2 chips because on an other place it looks like that there are only two chips supported at the moment.
Xmas and trinibwoy, can't believe I missed that. I was focusing on outdoor environments (Far Cry) and somehow ignored indoor environments (Doom 3).
It also makes sense that, given a high framerate, using the prior frame as a predictor of the next frame's workload would be pretty intelligent. But I'm still not sure this would eliminate the minimum framerate (the momentary dips upon encountering a new area or something), though it may reduce the duration of the dip (and thus raise the framerate of, say, the bottom quartile).
Did anyone benchmark AFR vs. SFR in the same game? AFR would seem to be the best solution, but I forget which mode has potential problems with certain effects and features. I guess it's already time to re-read Dave's SLI article. Edit: Apparently RTT affects both modes, so the ideal appears to be a "single" GPU composed of multiple cores (apologies to Chalnoth) that works with a single local memory pool, allowing for "full speed" spatial/temporal effects that may stretch across a frame split or alternating frames.
It also makes sense that, given a high framerate, using the prior frame as a predictor of the next frame's workload would be pretty intelligent. But I'm still not sure this would eliminate the minimum framerate (the momentary dips upon encountering a new area or something), though it may reduce the duration of the dip (and thus raise the framerate of, say, the bottom quartile).
Did anyone benchmark AFR vs. SFR in the same game? AFR would seem to be the best solution, but I forget which mode has potential problems with certain effects and features. I guess it's already time to re-read Dave's SLI article. Edit: Apparently RTT affects both modes, so the ideal appears to be a "single" GPU composed of multiple cores (apologies to Chalnoth
) that works with a single local memory pool, allowing for "full speed" spatial/temporal effects that may stretch across a frame split or alternating frames.This is not restricted to indoor scenes. Even in an outdoor environment you could have a rock, tree, etc. cover half of your screen while the other half shows a complex scenery. Or you could pass along a single complex building. And other effects, like multi-layer smoke trails... well, you should probably try to not have them centered in your viewPete said:
I think a horizontal split is better than a vertical split (both with dynamic adjustment) because most objects enter the field of view from the sides. Turning left or right is much more common than looking up or down.
Objects don't usually pop up in an instant, so even if one card suddenly has to do 10% more work, your framerate is still way higher than it is with a single chip setup.
Similar threads
