Sound_Card
Regular
And why is that?
So, do we know anything about RV670 yet?
- Thread starter nicolasb
- Start date
And why is that?
how are the chips going to get access to vram, the memory controller will have to change considerable to do what you are saying, which I think this was discussed in a thread along time back, it didnt' make much sense.
If you look at Fear numbers with AA and AF that Vertex Shader provided, the card is seems bandwidth limited, FEAR is one those games that is very scalable, but in this case, falls flat, double cards or chips doesn't double the available bandwidth, so thats a limitation that must be over come, if the bus size stays the same as we have now with 1 core, what is going to happen with 2 cores? Its not going to be 512 anymore, it will be 256 for each.
Sound_Card
Regular
You know for having almost half the bandwith, the x1950xt sure does not do that bad in FEAR.
http://www.xbitlabs.com/articles/video/display/radeon-hd-2900-games_6.html#sect0
http://www.xbitlabs.com/articles/video/display/radeon-hd-2900-games_6.html#sect0
actually it doesn't scale well at all for the pro
http://www.thetechlounge.com/article/304-5/ATI+Radeon+X1950+Pro+256MB+PCIE+Preview/
I think he is referring to the poor scaling in average fps and negative scaling in minimum fps in FEAR. The negative scaling is probably due to driver overhead in a CPU limited situation, but why does the average fps scale so poorly? IIRC NVidia does better in FEAR.
You are looking at very old drivers in that article.
true, but don't have any other data to go with 
http://www.rage3d.com/reviews/video/atix1950pro/index.php?p=4
here it does better without AA and AF, the scaling is much better, but with AA and AF scaling isn't so hot.
seems more to do with fillrates for FEAR, does fillrate double with crossfire or sli? Since one card is still rendering to the screen I mean, I understand that AA will be done 50/50 in AFR mode, but still rendering is always done by one card, the master card right?
http://www.rage3d.com/reviews/video/atix1950pro/index.php?p=4
here it does better without AA and AF, the scaling is much better, but with AA and AF scaling isn't so hot.
seems more to do with fillrates for FEAR, does fillrate double with crossfire or sli? Since one card is still rendering to the screen I mean, I understand that AA will be done 50/50 in AFR mode, but still rendering is always done by one card, the master card right?
Given that the only element that changes in the system in that test is either the addition of a second card or changing to a dual graphics card, the fact that peak FPS is scaling suggests that the limitation may be elsewhere.
I'm looking at data here that shows, with recent drivers, Crossfire scaling is up at the 1.8x level for more GPU limited situations in FEAR.
Sound_Card
Regular
Did you just completly ignore my link?
Jawed
Legend
My theory for R700 is that each GPU in the set is unified within a single "virtual memory address space" - so that GPU A can fetch directly from GPU B's memory, for example. Or from the L2 cache in GPU B, if the data is in that cache.
Hopefully this works well enough that AFR mode is not needed, and instead supertiling is used.
So I see a dual-die predecessor to R700 being a more "loosely coupled" version of this scenario, with the virtual-memory/cache architecture already in place.
Dunno what would differentiate the two, except perhaps it's easier to put in a fast inter-GPU link between two dies if they share a package, than if they share the main graphics card circuit board.
I'm thinking of Xenos in XB360, where the two dies have a 32GB/s link between them, and they are both mounted on the same package.
So a dual-R6xx GPU might have a slower link joining the two GPUs than R700 sees amongst its component dies.
Jawed
Hopefully this works well enough that AFR mode is not needed, and instead supertiling is used.
So I see a dual-die predecessor to R700 being a more "loosely coupled" version of this scenario, with the virtual-memory/cache architecture already in place.
Dunno what would differentiate the two, except perhaps it's easier to put in a fast inter-GPU link between two dies if they share a package, than if they share the main graphics card circuit board.
I'm thinking of Xenos in XB360, where the two dies have a 32GB/s link between them, and they are both mounted on the same package.
So a dual-R6xx GPU might have a slower link joining the two GPUs than R700 sees amongst its component dies.
Jawed
Sure you do, take a look at any more recent reviews from the 2000 series, such as this - we're talking about Crossfire scaling in this case and the same Crossfire hardware applies to X1950 PRO to the HD 2000 series, whats changed is the software. Bear in mind that since the eariler X1950 reviews came out the default Crossfire profile has changed from SuperTile/Scissor to AFR-Compatible, and, of course, we'll have also profiled more titles since then.true, but don't have any other data to go with
No we are talking about the pro not the XTX, the XTX is quite a bit more powerful of a card.
Dave, I know Crossfire scales very with the 2900 series, but lets say there is a heavy a bandwidth bottleneck or fillrate bottleneck, how would each of these effect a card in crossfire.
if we have a dualcore setup, with the current way of doing things, like we see in SLi or Crossfire, I don't see much of an advantage doing this, over a monolithic die, since the GPU's already has redundancy and ability to turn of parts as needed to get functional chips and this is the goal of going to multiple cores.
Sound_Card
Regular
My point was bandwith is not the bottle neck in FEAR.
Does not filrate scale with a second GPU?
http://www.elitebastards.com/cms/in...sk=view&id=387&Itemid=27&limit=1&limitstart=3
vertex_shader
Banned
I already explain in post #75 :smile:
SLI have this problem too, this is why i not see how can AMD/NV make competative product for the users with 2GPU-1PCB/2PCB, this cards only " we can't compete with a single GPU with try with 2 performance GPU" solutions, AMD problem when the card really come in 2008jan only its miss the in time release, as i write the possibility NV in 2008jan still have only 8800ultra as the highest perfomrnace card its not big, and i not see how 2xrv670 can compete a nextgen highend GPU even when its only a "G80" shrink with more SPU/TMU/ROP and higher clock speed.
Last edited by a moderator:
Again, Crossfire scaling is not card specific. When you have two graphics cards you have 2x the resources, and that applies to all graphics processing elements (the exception is vertex processing in SuperTile/Scissor)
But the data you pointed to doesn't really support it.
If that was the case you would always have 2 times the scaling, which isn't always the case.
Why is FEAR now having only 1.8 times the performance when using 2 cards over 1 why isn't it 2 times performance with 2 times the resources. On average this is pretty much what ends up with in almost all games around that 70-80% increase.
vertex_shader
Banned
Why not? HL2:E1 where single card already have great performance scaling very good, in FEAR scaling ~20% with lower minimum frames.
Generally speaking we characterise "1.9x" the maximum Crossfire scaling ratio - it can, theoretically be better, but there is so overall overhead added so 1.9x is the target. To me, something around the 1.8x range is deemed as "pretty good".
As I said, all the resources are doubled, however along with the overhead, what tends to be the gating factor in Crossfire/SLI systems is temporal data that needs to be shared between the GPU's for AFR rendering.
As I said earlier, the maximum framerates are scaling very well, which in this case can point to the limitation being elsewhere in the system. Add to the fact that we have other benchmarks that show FEAR scaling very well in GPU limited situations.
vertex_shader
Banned
There is possibility 2xrv670 based card GPU's is in the same die package?
