3dfx technology glossary/Spectre info

[PCResources]

Dave and Marty, i do know what double and triple buffering is about.

What i do not understand is how that would possibly solve the memory bandwidth problem, if you have a constant supply of information, running at an undetermined speed, supplying the memory with more constants that it can handle, outputing it to frambuffer and rendering it at maximum speed. With constant flow, this would mean that the output from the memory would be maximized by memory throughput, but this is not the case, as the chip itself is requesting data at higher speeds than what can be supplied it will be requesting blanks for some instances, and that is the problem.

uh oh, now i have probably said too much.

Patrick Palm

Am speaking for PC Resources

 

[DaveB3D]

I explained in a previous post.. look up higher..

 

[Wingznut]

Wow... Great thread, guys! (not sarcastic)

It's real nice (and unfortunately rare) to actually see video card technology discussed, without flames everywhere.

 

[Marty]

Dave, I believe that your idea may work. At least, it makes sense to me.

Marty

 

[Sephiroth_IX]

<< My thought was this: people are saying that when they enable v-sync you reduce the bugs. Now your thought is that there is a bandwidth limitation. Now if that were the case, there would be time when this isn't so much a problem (depends on the situation). So you'll have extra cycles that are being wasted. Use those to render to a second back buffer. That allows you to have a bit of a buffer for if a problem were to come up. >>

Well, this isnt so cut and dry. If you enable V-Sync, it will limit you at generally 85 fps (if running at the normal 85hz refresh.) Therefore, in a situation where the card is capable of doing 100fps, it will not be bandwidth limited. However, in a situation where there are more textures on screen, enough to bog it down to say 70fps, it could still cause anomolys as the memory bandwidth will be full from moving textures.

Am i off? I'm tryin to keep up

 

[Marty]

One thing I still don't understand is why the V5 would interrupt the rendering of a frame. Patrick, you say that it is the result of a memory bandwidth bottleneck. Why wouldn't it just take more time to render the specific frame? I think this is the fundamental question that must be answered.

Marty

 

[JayPatel]

i want one of these!!!!!!!!!!!

 

[Volenti]

Has anyone used the &quot;r_showtris&quot; command in Q3 to see if these drivers actually reduce the amount of hidden geometry that is rendered? or some other developer tool that would show that some kind HSR is taking place?

Untill we can actually see evidence of HSR (and not just skipping frames left right and crooked, which these drivers seem to do when benchmarking) then I feel that we are just pissing in the wind....

 

[Sunner]

Hey cool, PCResources its actually very interesting to read what you have to say once you lose the &quot;Im the king of the world&quot; attitude

And yes, that was a legit compliment.

 

[PCResources]

<< Well, this isnt so cut and dry. If you enable V-Sync, it will limit you at generally 85 fps (if running at the normal 85hz refresh.) Therefore, in a situation where the card is capable of doing 100fps, it will not be bandwidth limited. However, in a situation where there are more textures on screen, enough to bog it down to say 70fps, it could still cause anomolys as the memory bandwidth will be full from moving textures. >>



This is correct, the fps sometimes has little to do with the actual dataflow. When there is a lot of data to be processed the fps will go down but the dataflow will still be high, just like you said, that means that the memory will still have problems keeping up.

As i said before, and i meant it as a compliment, you are a smart one.



<< Why wouldn't it just take more time to render the specific frame? I think this is the fundamental question that must be answered. >>



That answer would mean going into specifics, wich is something that i cannot do, but if someone else figures it out, i can confirm it.



<< Hey cool, PCResources its actually very interesting to read what you have to say once you lose the &quot;Im the king of the world&quot; attitude >>



Thank you, i used to complain about everyone elses attitude, then i understood why everyone had that attitude towards me, so i lost mine, and now i am pleased to get some good answers instead of flames (sometimes someone calles me names, but now i know that it has nothing to do with my attitude towards them, rather their own lack of understanding, and when they know they cannot win the fight with intelligence, they try to win it by discrediting me...).

Sometimes i might be a little harsh on people, usually this happens when i try to tell them about something that i have been working on, something i know, and they just tell me how wrong i am, without any kind of explanation as to why they think so. That bugs me...

Patrick Palm

Am speaking for PC Resources

 

[Marty]

The only reason I see why the V5 would abandon the rendering of a frame while it is still incomplete is if it is trying to keep up with something else, which runs at a faster rate than the V5 is capable of. At lower resolutions, the V5 can keep up, which is why there are no distortions. At higher resolutions, however, it can no longer keep place, so it renders only incomplete frames.

Now, if you look at the benchmarks that some of the webpages have posted, there is virtually no performance drop going to high resolutions, in fact, there are increases part of the time. This tells me that the &quot;thing&quot; which the V5 is trying to &quot;keep up with&quot; is not affected in speed by resolution. Now my guess would be that the V5 is trying to keep up with the geometry data being sent over the AGP bus.

As the CPU performs the HSR, it sends only the visible geometry to the V5. If there is any kind of movement, new geometry needs to be sent to the V5. Now I have heard the V5 supports guardband clipping in hardware, and I take this to mean that standing at a point and simply looking around a little bit would not require the sending of new geometry. However, if you move, or the view changes significantly, the new geometry would be required.

Basically, the V5 is trying to keep up with geometry data that is being &quot;streamed&quot; to it from the CPU. At low resolutions, it can keep up. At high resolutions, however, the geometry data is arriving too fast for the card to keep up. It seems like the fresh geometry data has precedence over the old data, so when something new arrives, the old stuff gets thrown out and a new frame is rendered.

I wonder if anyone has a PCI V5 who could benchmark this. The slower bus may be a bottleneck to this streamed geometry data, resulting in lower performance.

Of course, I could be completely wrong.

Marty

 

[Varborta]

ok, i smell more intense competition coming from revenger 3dfx thats good

 

[Varborta]

Woowhoo!!!!! I just noticed I am a senior member now yes yes yes!!!

 

[Sephiroth_IX]

First off, excellent post Marty. That brings up another idea.


At high resolutions, there is much more video card activity. Whenever there is video card activity, the AGP bus is being used. Because the processor has to complete it's HSR functions before the frame buffer can be filled and thus the frame sent to the monitor. If there is a bottleneck in there, the card may still think it is completing it's functions and drawing a frame, thus not knowing it is sending a black one. This means it could be anything from system memory to AGP to processor speed that could be killing the frames.

 

[Soccerman]

this also makes me wonder, at higher resolutions, the V5 is both fillrate and bandwidth starved, yet it doesn't have these black frames.

it COULD try to keep up with the CPU, and result in black frames, however it DOESN'T so that is not the problem.

&quot;Because the processor has to complete it's HSR functions before the frame buffer can be filled and thus the frame sent to the monitor.&quot;

so you're saying that because the CPU timing is different, now that it has this extra work to do, it could be throwing the rest of the computer off, even though that doesn't happen when the bottleneck is the CPU in normal games?

for example, my computer is a K6-2 400 with Voodoo 3, with a crappy old monitor (max res is 640X480 with 60hz). I play Homeworld quite alot, and I know for sure that my bottleneck is the CPU.

yet, even though the CPU cannot send frames fast enough to keep up with the Video card, it does not introduce problems with the Video card thinking that it has recieved T&amp;L data and drawing a 'blank'.

it probably isn't the AGP bus either, becuase the end result of the CPU doing this is LESS data being sent through the AGP bus to the video card.

it could be a number of things if you ask me:
-a buggy HSR engine
-a feature that we don't know about causing problems

those are my personal likely suspects.. the drivers ARE the problem though, they control what the (somewhat limited) hardware does. if it is any of those two problems (video card not able to finish a frame quick enough, or CPU too slow and video card still draws without data), it would cause problems in everything else.

 

[Mr.IncrediblyBored]

Not very on topic, but still an amazing read over at Beyond3D about HSR and culling techniques in general.

Check this sucker out right here.

 

[DaveB3D]

linking B3D eh? sigh.. what would you guys do without me?

hehe

Edit: If you do some searching, you should be able to dig up plenty more threads on the subject. I know there are one or two where I go into a few ideas. Also, reference posts made by SA. He really knows his stuff on occlusion culling.

 

[Mr.IncrediblyBored]

<< If you do some searching, you should be able to dig up plenty more threads on the subject. I know there are one or two where I go into a few ideas. Also, reference posts made by SA. >>


How do you think I found this topic? I've seen several posts by him and decided to do a search.


<< He really knows his stuff on occlusion culling. >>


Shame I don't really understand a word he says. Still makes for great reading though.