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 .
Hmm, to reach 2T you can do

1280(alu's) *2(fma)*781MHz,
1600(alu's)*2(fma)*625MHz,

Since rv770 already runs at 750MHz, the new shder count seems to be 16 SIMD's, downright tepid compared to rv670->rv770 transition.

And then there is of course the question of what are they doing with the ~350mm2 die area?

The sentence was "more than 2 Teraflops", we remember that also the HD4850 was exactly 1 Teraflop, so this time it could be that HD5850=2 Teraflops and HD5870>2 Teraflops too. Moreover, to have a big enough difference in performance between 5850 and 5870 to justify the price difference (as both will almost surely use GDDR5 this time) I think ATI has two options:

- to lower clock of 5850 substantially (20% or more)
- to make the 5850 a lower SIMD-count chip

Both options hints to an high amount of SIMD, as if the first one is true, then HD5850 needs an high number of SP to reach 2 teraflops, if the second is true, then it means that HD5870 will have a lot more SP than 5850 -> the maximum count of SP on the chip is high. I seriously doubt we'll see 900+MHz parts at launch, considering that tha 40nm process seems to have still some problems and that there could be thermal issues.
If we make a very rough calculation, a 2xRV770 chip in 40 nm will measure about 275 mm^2, let's throw in a 20% more area for DX11 compliance and improvements, and we have 330 mm^2. So if 350 mm^2 for Cypress is true, then it could fit well with a 1600 SP, 8 RBE chip (I hoped for evem more, though, as my first guess was 2000 SP, but who knows).
 
I think we have a view... If we consider RV830 / 180mm² to be 640SPs part, then a GPU, which has almost double die-area (so likely twice as much transistors, +90-95%), shouldn't be based on 1280SPs.

RV350 - R350: 2x ROPs/TMUs/ALUs/bus => +50%
RV410 - R420: 2x ROPs/TMUs/ALUs/bus => +33%
RV530 - RV570: 3x ROPs/TMUs/ALUs, 2x bus => +120%
RV630 - RV670: 4x ROPs, 2.7x ALUs, 2x TMUs/bus => +70%
RV730 - RV740: 2x ROPs/ALUs, + GDDR5, + DP support => +60%
RV730 - RV770: 2x ROPs, 2.5x ALUs, 2x bus (+GDDR5), +DP support => +85%

I'd say with +50-60% (280-290mm²) doubling of almost every functional part of the GPU should be possible. For a 90% bigger GPU I'd expect something like 2x ROPs/bus, 2.5x ALUs/TMUs.

If RV830 has 640 SPs, than 1600 SPs (2.5x) seems to be in the range of possible configurations for RV870 (350mm²). If RV830 will have 800 SPs, I wouldn't be surprised by 2000 SPs for ~350mm² GPU.

Yes, we don't know how costly the DX11 implementation is or how many SPs does the RV830 have. But we know size ratio of two dies, so we can base the estimation upon proporcional difference.

While I respect your efforts above you're not telling me anything different to what I said. With so many "ifs" and conditionals you've used there you couldn't answer that question even if you'd knew what kind of die area X11 requirements would consume.
 
My point was that the increase in flops (and the associated increase in tex, rops, alu,tex,rop ratio assumed static of course) is too low to fill up 350 mm2 worth of area. I can only assume that it must be taking up a lot of area, because otherwise we'd prolly have seen a larger increase in alu's and texturing.


IMHO, I'm still not sure about that. AMD is said to present a few GPUs together, RV830 and RV870. So maybe (just maybe) the 2TFlops is the Flops-rating of the RV830 and not the RV870. AMD has played this game before with the RV770 launch but with wrong informations. Maybe now they do the same, giving out RV830 informations but saying it is RV870 information. :)
 
To cut costs, it'd make sense, to first market the fully functional and salvage parts (i.e. xx50 and xx30) and later add the speed binned xx70 model - given their supposed speed increase over the competition, DX11 not to mention. Those two factors should provide enough momentum for great sales.
 
IMHO, I'm still not sure about that. AMD is said to present a few GPUs together, RV830 and RV870. So maybe (just maybe) the 2TFlops is the Flops-rating of the RV830 and not the RV870. AMD has played this game before with the RV770 launch but with wrong informations. Maybe now they do the same, giving out RV830 informations but saying it is RV870 information. :)

Well, nv is surely not going to launch before sep 10, soft or hard, so there is no downside of letting the fud machine stop. But then, why not screw with people's minds and enjoy the fun...?:rolleyes:

There are less than 2 weeks to go now, we shud really be seeing some real things pop out.

Assuming rv830 is running at 2T, we have

640 (alu's) * 2 (fma) * 1.56GHz
800 (alu's) * 2 (fma) * 1.25GHz

:oops:

So definitely, the 2T chip is a 5850/5870 part. Now 2T could very well be 2-2.3 ish, but I am sure that the part referred to there, is definitely not more than that.
 
Back
Top