AMD: R8xx Speculation

How soon will Nvidia respond with GT300 to upcoming ATI-RV870 lineup GPUs

  • Within 1 or 2 weeks

    Votes: 1 0.6%
  • Within a month

    Votes: 5 3.2%
  • Within couple months

    Votes: 28 18.1%
  • Very late this year

    Votes: 52 33.5%
  • Not until next year

    Votes: 69 44.5%

  • Total voters
    155
  • Poll closed .
From the CAT 9.8 thread, release notes would appear to indicate that ATI engineers have been doing a bunch of work on 4 GPU mode Crossfire building up to this launch :cool:
http://www.pcper.com/comments.php?nid=7641
The multi-threading features of D3D11's driver model are partly about allowing the CPU to spread its driver bottlenecks around cores.

X3/X4 configurations won't scale in cases where there's a CPU bottleneck, so those huge scalings could be as a result of a significant re-working of the drivers to better use the CPU.

Jawed
 
I hope the port cluster isn't just some ES board anomaly, it's time for some DisplayPort goodness and the chicken has to lay the egg here. The monitor manufacturers might care if cards actually start using the standard.
 
Pretty sure this is wrong (though I don't know if that was also true for the older rs690 IGP). Sideport memory is only 16bit. At least I haven't seen any board which uses 2 memory chips...
Looking here : AMD 700 Chipset Series

Claims the RS780(780G mainstream), RS780D(790GX performance) both support DDR2 and GDDR3. Also RS780E(780E embedded) supports GDDR3.

From anandtechs preview of the 785G(RS880) this image also shows sideport DDR3 support on the gigabyte box.

Quickly looking couldn't find anything useful on the AMD site itself saying one way or the other. :cry:
 
In theory RV740 and these Evergreen GPUs were in development long before AMD found out that RV770 had enough space to fit in 2 extra clusters (i.e. 10 instead of 8).

I'm thinking that both RV740 and Juniper might have the same cluster count, i.e. 8, as they're both derived from the original RV770.

In general with the higher performance ATI cards the texture count matches up with the interpolators, which are effectively defined by the rasteriser, which is effectively defined by the RBE count. i.e. 16 RBEs means 32x interpolations per clock, matches up with 32 TUs. RV770 has more texturing than interpolation, which is the reason for some funny numbers in texturing synthetics.

So, RV740 came out with 16 RBEs and is basically the original specification for RV770 scaled to 40nm (presumably without sideport, too).

I'm now thinking that Juniper is RV740 with D3D11 added on, but otherwise carries the same number of units: 16 RBEs, 8 clusters, 640 ALU lanes, 32 TUs. This is what happened with R600->RV670, D3D10->10.1, the memory interface was cut in half, but the unit counts remained the same at the same time as jumping a process node. Though, of course, D3D10.1 hasn't been seen as a major change in design for ATI, generally speaking.

Still, the idea that there's >50mm² of D3D11-specific changes is kinda worrying. OK, so there might be 5 or 10% of that which is purely performance tuning, but overall it seems like a big hit.

Jawed
 
Looking here : AMD 700 Chipset Series

Claims the RS780(780G mainstream), RS780D(790GX performance) both support DDR2 and GDDR3. Also RS780E(780E embedded) supports GDDR3.
This cannot be true, since we know it has to support DDR3 (all rs780 boards I've ever seen with sideport memory had a 16-bit ddr3 configuration). I think someone just mistook ddr3 for gddr3...

Quickly looking couldn't find anything useful on the AMD site itself saying one way or the other. :cry:
There's unfortunately no datasheet available for rs780. There is one for rs690, but for sideport it refers to another (not public) doc...
I found this though, http://global.hkepc.com/747 which seems to confirm rs780 is indeed 16-bit ddr2/ddr3.
 
Jawed- Just for clarification when you say Juniper are you talking about the RV830/181mm2 GPU?

While your logic does make sense, I think a shrunk RV770 would still be possible.
 
I mean 2nd generation of DX11 graphics cards, cause they use 32/28nm process from TSMC, not 40nm.

Yeah, cause we all KNOW how reliable TSMC roadmaps are.

so lets see, TSMC says 2010 so add 12-18 months and you'll be on target. They are just now getting their 40nm process to the point that it can actually run.
 
When a corporation wants to consider themselves as a serious corporation, they owe to provide solid support for everyone. Categorising and prioritising their clientelle's needs is perfectly understandable but i am not talking about an issue that jumped up yesterday!

Maybe they have prioritized the issue and you just don't agree with their priority. Honestly, in the PC space, if a TV is advertising 24 Hz, that's probably what you are going to get. Maybe the problem lies not in their drivers but in your TV?
 
I'm thinking that both RV740 and Juniper might have the same cluster count, i.e. 8, as they're both derived from the original RV770.

So, RV740 came out with 16 RBEs and is basically the original specification for RV770 scaled to 40nm (presumably without sideport, too).

I'm now thinking that Juniper is RV740 with D3D11 added on, but otherwise carries the same number of units: 16 RBEs, 8 clusters, 640 ALU lanes, 32 TUs. This is what happened with R600->RV670, D3D10->10.1, the memory interface was cut in half, but the unit counts remained the same at the same time as jumping a process node. Though, of course, D3D10.1 hasn't been seen as a major change in design for ATI, generally speaking.

Still, the idea that there's >50mm² of D3D11-specific changes is kinda worrying. OK, so there might be 5 or 10% of that which is purely performance tuning, but overall it seems like a big hit.

Jawed

IMHO, you forgot that you need a lot of Die-area to handle all the fixed units which don't change between GPUs so much. When I had made a comparison of all the 55nm ATi GPUs ( RV630, RV730, RV670, RV770) the spec (amount of shaders and TMUs) conformed to the Die-area only when I subtraced roughly 75-85mm² from the Die area. So maybe for 40nm you have to subtract maybe 50-60mm² from the die area to get the net area available for shaders and TMUs. Therefore the RV740 would have 77-87mm² of area available for shaders and TMUs, and the Juniper would have roughly 120-130mm² available for shaders and TMUs.

Therefore I wouldn't be surprised if the Juniper has a much higher shader count and TMU count as you seem to expect.
 
The new FirePRO cards all have DP support from the get -go. Isn't too far to speculate that evergreens will have it too.
 
Back
Top