AMD: R7xx Speculation

Status
Not open for further replies.
A chip so big and fast that we don't need 2 of them on a single board? :oops:
Pretty much. Look at it this way: a dual-chip config couldn't go above a 300W TDP, and that's already pushing it. So what's I'm effectively proposing is a 200-225W single-chip. And I'm just guessing here, I have zero insider info that'd point in that direction (but neither do I have any that'd really point at it being false).
 
Pretty much. Look at it this way: a dual-chip config couldn't go above a 300W TDP, and that's already pushing it. So what's I'm effectively proposing is a 200-225W single-chip. And I'm just guessing here, I have zero insider info that'd point in that direction (but neither do I have any that'd really point at it being false).

It would certainly be a surprise similar to the whole unified/non-unified adventure with the G80 before launch. But one might question what the purpose of the 3870X2 was, if that is the case?Why use up resources on board design, drivers and so on, for such a short-lived approach?
 
Pretty much. Look at it this way: a dual-chip config couldn't go above a 300W TDP, and that's already pushing it. So what's I'm effectively proposing is a 200-225W single-chip. And I'm just guessing here, I have zero insider info that'd point in that direction (but neither do I have any that'd really point at it being false).
I completely forgot about the power ceiling.

http://www.techreport.com/articles.x/14284/8

Comparing load power consumption, 3870's 222W against X3's (3870+X2) 453W we get 231W for X2 (though that assumes that X3 is running at full draw which seems unlikely due to the vagaries of scaling).

Jawed
 
Yes, and?What's the precise limiting factor in that?
It's precisely the exact difference between the time taken to transmit any kind of data using either PCIe-1.1- or PCIe-2.0-bandwidth. I don't know about latencies there, but raw bandwidth seems to double AFAICT from the specs.

It is my understanding, that even in X2-configuration there still is a primary GPU and a secondary one, the latter of which has to transmit it's finalized buffers (and more, if there's some evil post-processing at work) to the framebuffer of the first one. Transmitting-time only should be halved thus.
 
Given these rather realistic measurements over at xbitlabs.com (http://www.xbitlabs.com/articles/video/display/gainward-bliss9600gt-512gs_7.html#sect0), I'm having a hard time putting the X2 significantly above 170 Watts with this specific workload.

Xbit has tested an HD3870 512 with 775 MHz, HD3870X2 is clocked at 825 MHz per GPU.
Well, their numbers are lower than everyone else since it doesn't consider the PSU's ~80% inefficiency. But anyway, the point is that the cooling requirements would be too high with a >300W TDP...
Question: Does a videocard consume during 3Dmark 2006 at 1600x1200&4xAA+AF 16 more than during UT3 at 2560x1600 (TR)?
I hope, I'm not wrong, but TR's HD3870 was clocked with 775 MHz.
 
Last edited by a moderator:
Well, their numbers are lower than everyone else since it doesn't consider the PSU's ~80% inefficiency. But anyway, the point is that the cooling requirements would be too high with a >300W TDP...
Right, and xbit does measure GPU only without the rest of the system - which hardly anybody else does.

I agree, standard dual-slot air cooling wouldn't be possible for such power levels at any bearable acoustics. But maybe some clever tricks are for this kind of setup are yet to be uncovered? The X3-card for example has at least one of it's MXM-modules on the respective other side than the two other GPUs. Tri-Slot-cooling sounds ridiculous - but two power connectors and the like did that too some time ago.

I'm not saying that this will be the case, but i wouldn't rule this possibility out altogether, given the sometimes quite desperate measures to get the fastest benchmark bars.
 
Xbit has tested an HD3870 512 with 775 MHz, HD3870X2 is clocked at 825 MHz per GPU.
That's an almost 6,5 percent difference... plus the different GDDR-types, plus the bridge chip, plus - and in my book the most important difference - on X2 you don't have the doubled efficiency losses in power regulation circuitry compared to dual 3870s.

That's exactly why i was calculating some headroom of about 170-(81,1*2) watts.
 
It's precisely the exact difference between the time taken to transmit any kind of data using either PCIe-1.1- or PCIe-2.0-bandwidth. I don't know about latencies there, but raw bandwidth seems to double AFAICT from the specs.

It is my understanding, that even in X2-configuration there still is a primary GPU and a secondary one, the latter of which has to transmit it's finalized buffers (and more, if there's some evil post-processing at work) to the framebuffer of the first one. Transmitting-time only should be halved thus.

Umm, that would be correct if data would be transmitted through the PCIE switch and not a high-speed interconnect. Just like in normal CF, the cards "talk" through the CF connector(s), not through PCIE. Without attaching the CF connectors you're indeed bound by PCIE performance, but you're also functioning in a worst case scenario.

Question:why do you think there is only 1 CF connector on 3870X2(asides from the obvious lack of support for more that 4 GPU configurations?)?Another question:why does nV still use an internal SLi connector on the GX2, in spite of them already having a PCIE2.0 switch?;)

The PCIE switch itself is not a limiting factor on the X2, in spite of it being only 1.1. The difference moving to a 2.0 one would bring wouldn't be worth the cost at all, and would be borderline undetectable.
 
Pretty much. Look at it this way: a dual-chip config couldn't go above a 300W TDP, and that's already pushing it. So what's I'm effectively proposing is a 200-225W single-chip. And I'm just guessing here, I have zero insider info that'd point in that direction (but neither do I have any that'd really point at it being false).

"R"T200? :D

If we go further on this thought I'd propose a 180-200W chip with the diesize of pre-G80 peers. And that's with high clocks. >200 would be a challenge to cool in a limited diespace.

But again, there's the nagging rumor of the notebook ATi flagship. ;)
 
It would certainly be a surprise similar to the whole unified/non-unified adventure with the G80 before launch. But one might question what the purpose of the 3870X2 was, if that is the case?Why use up resources on board design, drivers and so on, for such a short-lived approach?

What makes the drivers so much different than a pair of RV670's?:LOL: Not really much.

The life is just fine. The card deputed in January and suppose RV770 hits in summer, thats 6-7-8 months on the market.

On the whole x2 and monolithic talk, I would think their would be a x2 of some sort in 2009 with 45 or 40nm shrinks. As far as RV770 being a monolithic beast, I say it's pretty possible.
 
What makes the drivers so much different than a pair of RV670's?:LOL: Not really much.

The life is just fine. The card deputed in January and suppose RV770 hits in summer, thats 6-7-8 months on the market.

On the whole x2 and monolithic talk, I would think their would be a x2 of some sort in 2009 with 45 or 40nm shrinks. As far as RV770 being a monolithic beast, I say it's pretty possible.

The difference is in terms of focus:whereas CF wasn't exactly a top priority before, now with an X2 type of card it became one. So resources were invested in tweaking scalability, AFR compatibility etc.

The RV770 itself is almost certainly a single chip solution, I think Arun was speculating about the R700 card. Which if things remain as they are now, should be a dual RV770 solution. If that's not the case...cool, I'm just as curious as everyone else to see how things actually turn out.
 
Umm, that would be correct if data would be transmitted through the PCIE switch and not a high-speed interconnect. Just like in normal CF, the cards "talk" through the CF connector(s), not through PCIE. Without attaching the CF connectors you're indeed bound by PCIE performance, but you're also functioning in a worst case scenario.
I was and am under the impression, that exactly this switch manages data transfer between the two chips. Do you have anything more 'official' than your impression that this is not the case? I'd be greatly interested in that.

Question:why do you think there is only 1 CF connector on 3870X2(asides from the obvious lack of support for more that 4 GPU configurations?)?
Why would there be more connectors when only coupling two cards? As seen in single-gpu quad-CF, radeon cards are not necessarily all interconnected as is the case with tri-SLI, so one connector should be sufficient.

Another question:why does nV still use an internal SLi connector on the GX2, in spite of them already having a PCIE2.0 switch?;)
Maybe because the two boards of a GX2 do not share other physical connections besides some screws and the cooler - and i'd rather doubt, nvidia does transmit data via either of those. ;)

The PCIE switch itself is not a limiting factor on the X2, in spite of it being only 1.1. The difference moving to a 2.0 one would bring wouldn't be worth the cost at all, and would be borderline undetectable.
Well, if you're correct and there's a hidden CF-link somewhere on or in the PCB, I'd agree. But since more than once 'real' CF on a PCIe-2.0-Board is faster than X2, I not totally convinced to attribute that only to faster memory clocks on the former - not taking into account the higher engine clock on the latter.
 
To be blunt: it would seem silly if X2 connects two RV670s solely through the PCI Express interface (via PLX chip) - after all each chip has two CrossFire interfaces waiting to be used.

Jawed
 
So it would seem. But it does not explain, why two single HD3870 in CF are sometimes (significantly) faster than the X2.

I'm not saying, they're using the PCIe 1.1 exclusively but as of now I'm not convinced of the opposite.
 
The CrossFire connections between boards is for the framebuffer data to be passed for compsoting by the master card. Inter-frame communication (i.e. off screen data that needs to be pass from one GPU to the other) is handled by PCIe. Both elements are hardwired in an X2, although X2's have the additional element that the switch can pass data directly to each of the GPU's, where two boards in CrossFire have to communicate via the PCIe controllers.
 
So it would seem. But it does not explain, why two single HD3870 in CF are sometimes (significantly) faster than the X2.

I'm not saying, they're using the PCIe 1.1 exclusively but as of now I'm not convinced of the opposite.

That is probably a consequence of the significant bandwidth difference between the two configurations you mentioned.
 
The CrossFire connections between boards is for the framebuffer data to be passed for compsoting by the master card. Inter-frame communication (i.e. off screen data that needs to be pass from one GPU to the other) is handled by PCIe. Both elements are hardwired in an X2, although X2's have the additional element that the switch can pass data directly to each of the GPU's, where two boards in CrossFire have to communicate via the PCIe controllers.

Isn't the bandwidth requirement for the inter-frame stuff higher than the framebuffer copy? It seems strange that there would be a dedicated connection just for that.
 
The CrossFire connections between boards is for the framebuffer data to be passed for compsoting by the master card. Inter-frame communication (i.e. off screen data that needs to be pass from one GPU to the other) is handled by PCIe. Both elements are hardwired in an X2, although X2's have the additional element that the switch can pass data directly to each of the GPU's, where two boards in CrossFire have to communicate via the PCIe controllers.
Thanks for the explanation, Dave! Can you tell us more what you mean by "hardwired"? What use is the PCIe-switch in X2 at all?

That is probably a consequence of the significant bandwidth difference between the two configurations you mentioned.
Yes, possibly. But the HD38x0 are not known for being bandwidth starved.
 
Status
Not open for further replies.
Back
Top