ATi 4870X2(R700) PCB Debut

[Shintai]

Why post so much for a PCIe 2.0 switch chip? It could be godly and nothing changes for the texture loading and such from CPU and memory. Still 1 PCIe x16.

And at best, PCIe x16 2.0 is 8GB/sec in each direction.

Also the slides above just tells its easier to assign buffers and such if you partition the ports in a whole other faction with alot of other devices. And not just a static 3x x16. It could also be 12 x4. Thats why the buffers are as shown.

Also it has to be a 8648 chip, with another 140ns lag added.

http://www.plxtech.com/products/expresslane/pex8648.asp
http://www.plxtech.com/pdf/product_b...95_16Jun08.pdf

[hollo]

this :D?

[Hornet331]

Why post so much for a PCIe 2.0 switch chip? It could be godly and nothing changes for the texture loading and such from CPU and memory. Still 1 PCIe x16.

And at best, PCIe x16 2.0 is 8GB/sec in each direction.

Also the slides above just tells its easier to assign buffers and such if you partition the ports in a whole other faction with alot of other devices. And not just a static 3x x16. It could also be 12 x4. Thats why the buffers are as shown.

Also it has to be a 8648 chip, with another 140ns lag added.

http://www.plxtech.com/products/expresslane/pex8648.asp
http://www.plxtech.com/pdf/product_b...95_16Jun08.pdf

what surprises me is, that the pcie 2.0 switch has a higher latency then the pcie 1.0 switch.

[Helmore]

this ?

That would mean that it is in fact just another version of R680, just with the RV670 chips replaced with the RV770 chips and of course a different bridge chip.
My guess would be that your "standard crossfire bridge in PCB" would be the "CrossFireX SidePort" which would in fact different than the normal bridge....or the CrossFireX Sideport is another parallel connection.
Just speculation though, and hoping....

[Frank M]

He does suggest that it is not yet listed, not much on how custom the chip really is....
Here is hoping that the CrossFireX Sideport is a separate interconnect between the 2 GPUs, a (more) direct connection between the GPUs as this will cut latencies and may increase bandwidth.

Well, I think them changing the good and efficient RingBus for the hub-system
happened for a reason... and I think that reason is shared memory between
the gpu chips.

[Tonucci]

No idea. Had no interest in ATI untill now. I know that two 9800GTX's in SLI worked better then a single GX2.

When clocked the same the GX2 is an hair better, due to decreased latency IIRC.

[Helmore]

Well, I think them changing the good and efficient RingBus for the hub-system
happened for a reason... and I think that reason is shared memory between
the gpu chips.

I'm pretty sure that won't happen and then there is another fact and that is that this memory subsystem they are using on the RV770 cards is more efficient than the ring-bus. The Ring-bus had other reasons for it's existence.

And then of course the fact that there won't be a shared memory system between the GPU chips, there will probably be some seriously improved communication but no complete memory sharing. Not this generation at least and probably not even the next generation.

[Helmore]

When clocked the same the GX2 is an hair better, due to decreased latency IIRC.

There is also the fact that R700 will be using 2 RV770 chips clocked at 4850 speeds as the cooling needed to do 2 of those at 750 MHz. is just too crazy. This will mean that R700 will be faster than 2 4870 in CFX in general, but in occasions where crossfire scales exceptionally well this will be different.

BTW, keep in mind that I'm just speculating although I'm pretty sure on those clocks.

[annihilat0r]

Ouch, is it confirmed that R700's clocks will be around 625?

[Unbornchild]

that R700 might be faster than 2 4870 in CFX in general

you're only speculating, so....

I will buy 2 4870's anyway because i won't be able to resist. But probably i'll buy a 4870X2 later as well, and move the 2 single cards to my other PC. Then i finally will see the difference (if there is) with my own eyes.

Ouch, is it confirmed that R700's clocks will be around 625?

Someone who posts here regularily and knows, said that, yes. It is done because of the cooling factor. so if you want to clock them higher, change the cooler to some ultra-turbo-superdidooper one.

[nsidious]

definately a good idea to go with a black pcb

[adamsleath]

Then i finally will see the difference (if there is) with my own eyes.

and the difference will be marginal. takes no nostradamus to say that.

[G0ldBr1ck]

Relax and take a deep breath, The AMD folks have been leaking little tidbits of info out on this for awile now that makes it fairly obvious that the GPU interconect is not Xfire dependent at all but a PCI bridg chip still is needed for both GPU's to be fed at same time by the MB. The R700 will have to outperform 2 4870's wich really completly on the pci express bus. AMD has said many times that the interconnect tech. is superior to the R680 and I have no reason to doubt it. I Dont realy understand why so many are insisting that becouse it has a PLX Switch that its exactly like the 3870x2. There is no way at this day and time for them to do this without a switch, but that dose not mean it is the only link between GPU's.


Edit: actualy Im not sure that 2 4870's really completely on pcie bus There may be advances made here as well, but it has to be more dependent on it than R700

[Shintai]

The R700 will have to outperform 2 4870's

Why? It only needs to outperform one. And as said by others. Its most like 4850 clocks with GDDR5. Else the TDP would go completely skyhigh.

4850 is already the same TDP as 3870. An X2 with 4870 GPUs would be 50% higher TDP or so than a 3870X2...

[Unbornchild]

so many are insisting that becouse it has a PLX Switch that its exactly like the 3870x2.

Not many, only a few nv-fanboys (who don't feel very good lately) keep persisting, as if they know anything...

[G0ldBr1ck]

I would speculate that the interconnect on the X2 will out way the slower clocks vs 2 4870's

[G0ldBr1ck]

Not many, only a few nv-fanboys (who don't feel very good lately) keep persisting, as if they know anything...

Maybe they want to see AMD/ATI go under so that they can pay $2000 dollors for there Intel quad and $1500 for there NV Card

[LordEC911]

what surprises me is, that the pcie 2.0 switch has a higher latency then the pcie 1.0 switch.

Why does that surprise you?
Look at DDR...

[annihilat0r]

ATI really needs to have done a good amount of improvement to its x2 setup for the 4870x2 at 4850 clocks to be faster than two 4870s in CF.

[doompc]

Anand says the new Crossfire Sideport (wich is suposed to enable the GPUs to share memory) isn't enable in single RV770, and will be enable in the X2.
http://anandtech.com/video/showdoc.aspx?i=3341&p=9

And unlike the 3870X2, the new one has GDDR5 memory (probably faster than on normal 4870) and PCIe Gen2 interconnect.
Add binned, higher clocked GPUs and we have real competition to the GTX 280.

And I read somewhere the 4870X2 will have a full copper + heatpipe cooler, maybe with the fan in the middle, like the 7900GTX cooler.
I hope they bring back the "wind tunnel" blower design, like the X800/X1800 cooler.

BTW, I see a lot of CPU design influence here:

[naokaji]

Why? It only needs to outperform one. And as said by others. Its most like 4850 clocks with GDDR5. Else the TDP would go completely skyhigh.

4850 is already the same TDP as 3870. An X2 with 4870 GPUs would be 50% higher TDP or so than a 3870X2...

binned chips at a lower voltage maybe? dunno, just guessing, but its the only other way I can think off how they could keep heatoutput within reasonable levels.

[hollo]

ok, so according to anandtech it's more like this-

[zerazax]

ok, so according to anandtech it's more like this-

That seems likeliest. The sideport probably communicates certain information to improve scaling and performance while the bridge chip is used to communicate with the PCI-E interface (which is necessary). So it won't be CF on the chip but a different method.

[G0ldBr1ck]

http://images.anandtech.com/reviews/video/ATI/4800/hub.png[/img]

this photo shows the card using the same cooling design as 3870x2 or very similar.

[Helmore]

More like this:

Although it has been said by some that even this cooling design has been revised to improve cooling, there may be an extra heatpipe for example.