R700 Inter-GPU Connection Discussion

It also mentions the releasedate and new Vantage Extreme scores.

 

August 12th it is then

And 6920 in Vantage Extreme... how does that compare to a GTX 280?

 

2.4TFLOP / 8.8GFLOP/Watt = 270W. Scorchy.

 

Not good.

 

(DAAMIT!)

I received my 4870 this week and the continuous fan speed change is getting very annoying...

IT would be nice if someone (catalystmaker?) could shed some light on this!

Thanks!

 

ZOMG!!
Micro-stuttering is dead on R700, according to this source:

 

I found one source that claimed 4800.

And that source also claimed 5500 on R700..so 6920 is a huge improvement on that.

 

Actually, it's quite misunderstood.

It will not have the "Notebook" part of Powerplay- aka idle off. It will throttle, but not die off awaiting an initiation invitation from the SMBus.

Otherwise it should work the same. Hopefully that is.

 

4800 is about right without PhysX and a not-so ub0r-CPU (read: only dualcore) behind it.

 

Interesting, I wonder if they've done anything to address input "lag."

So it's rumored to be cheaper than GTX 280. Best case when CF is working is 180% the speed of GTX 280, average is 150% the speed fo GTX 280...and when CF completely falls on its face, then you have a 4870. Which in some cases is as fast as or faster than GTX 280 (in rare cases) and on average within spitting distance. And in worst case not so much.

Definitely shaping up to be an interesting summer. I wonder what Nvidia's response to this will be. They've ALWAYS had some kind of response in the past.

Regards,
SB

 

That response would definitely not be a pair of G200's in a bricked sandwich.

 

Cards mounted in 5.25" drive bays with cables running to the PCI-e slots? Even better if they just made them external. Then they could be used to heat your coffee.

To truly solve micro-stuttering doesn't that imply they're using some method other than AFR to render the scene? The only other way would seem to be to force a framerate (ie vsync) but that would apply to any crossfire setup I'd assume.

If they really wanted a multi-chip setup to work properly I still think we're back to the shared memory setup. I wouldn't be surprised if all the current cards could do it but they're just lacking bandwidth and hurting for latency. Might explain the high bandwidth utilization problems that have cropped up lately.

 

IF that's true (and applies to non-profiled games too) then ATi just won an internets.

MS bothers me so much with my 2x8800GTXs that I simply turn one card off.

 

Why exactly is screenspace tiling, say in an upper 60 percent and a lower 40 percent, such an irrelevant option? Especially if all your GPUs have a massive advantage over the competition in Setup-rate/throughput?

 

Both chips would (afaik anyway) need to do full geometry for every frame instead of every 2nd frame like with AFR

 

I understand this would be true for the now discontinued Supertiling-mode. But is this also the case even if you divide the screen into fixed stripes? Vertices should have positioning info about where on screen they like to be drawn after all.

Of course, you'd need some driver work and maybe some overlap so that you have to render 105 percent of the actual screen-res height to compensate for geometry not sticking to it's half.

 

And what happens when say, the lower part of the screen has 500% more work than the upper part?

 

=>Kaotik: That shouldn't happen. If I'm not mistaken, CrossFire SFR uses a fixed ratio for each game - but it's not always 50:50 'cause it's a known problem that there's more work in the lower part of the screen. And AFAIK, SLI continually shifts the ratio so the workloads are balanced better, so this should be possible with CF as well, but probably has more (CPU?) overhead.

 

Yes but in those cases both cards/chips make the same geometry calculations for every frame, which was something that CarstenS apparently at least hoped could be possible to be bypassed by making the dividing ratio fixed instead.

 

Thing is you don't know the final screen space position of a vertex until the vertex shader (and geometry shader) has been run. Each of the cards will have to run through all that data before they can cull the stuff that's outside their tile.