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 .
Sigh, why didn't I think about chopping up an RV770 die picture to make something 2x, instead of making a boring spreadsheet?

Jawed
 
A. It doesn't look symmetrical to me. Anybody?

B. It doesn't seem like photoshopped from rv770 die shots

C. rv740's die shots weren't leaked since after it's launch. So I can't say if it is authentic. What do old time gpu followers have to say about this?

D. Bottomline, how realistic/credible do you think the die shot is?
 
A lot of units of the same type and structure are grouped on the same half of the chip.
The GPU wafer that was photographed before whosed a quadrant-type arrangement that would look very different if it were grouped like that photo.

If this isn't the same chip, I would wonder about the need to change the overall structure between chips in the family.
 
AFAIK, SLI in SFR automatically adjusts the splitting line to balance loads. But it's still slower than AFR, probably an inherent SFR problem.

SFR will pretty much always be "slower" than AFR. After all, with AFR you are just basically doing the same thing twice. Latency of the frame stays the same with AFR but FPS "improves" because for each frame you are actually rendering two frames.

SFR OTOH, will actually reduce frame latency. The primary bottlenecks of SFR are tri rejection and shared resource access (aka render to texture, etc). Tri rejection is a bounding box problem can can be solved with a little overhead which basically means you don't get any speed up in the front end of the pipeline. The shared resource issue can be solved via inter-chip bandwidth. In general though, SFR is the better path to go down as it should result in more "noticeable"/"real" performance than the hack that is AFR.
 
http://img35.imageshack.us/img35/3797/rv870.jpg

bad PS work... RV770 die-shot + some units (the fuzzy part) taken from a die-shot of something else

That's what it looks like to me as well. Although the center portion is out of focus it looks a tad more zoomed in then the rest of the photo. And the edge of it looks cut off. The center of that photo is too well defined. You can clearly see a rectangle (albeit elongated) which doesn't seem to fit the rest of that photo.
 
http://img35.imageshack.us/img35/3797/rv870.jpg

bad PS work... RV770 die-shot + some units (the fuzzy part) taken from a die-shot of something else
The row of memory interface pads at the top is cropped from the Phenom die-shot, as well as the middle block thingy (which is from the HyperTransport hub). :LOL:

70613114.png
 
Last edited by a moderator:
You can add a sequence number to tris and not rasterize one till all previous ones have started rasterization (easier said than done). This way you can parallize the vertex pipeline and setup between the cards, so if you just use tiling to balance the pixel load it becomes easy to maintain balance (even relatively large tiles like say 512x512 sub-pixel samples should be enough).

Didn't ATI try this originally with their super tiling or whatever they called it? They split the screen into a checkboard pattern and had each card render the odd or even tiles.

Although I haven't got a clue whether they tried to add sequence numbers to triangles or not.

Tiling would appear to alleviate any sort of uneven workload, but I'd imagine it's quite a bit more difficult to implement than a simple split frame.

Regards,
SB
 
Didn't ATI try this originally with their super tiling or whatever they called it? They split the screen into a checkboard pattern and had each card render the odd or even tiles.

Although I haven't got a clue whether they tried to add sequence numbers to triangles or not.

Tiling would appear to alleviate any sort of uneven workload, but I'd imagine it's quite a bit more difficult to implement than a simple split frame.

Regards,
SB

Supertiling was also supposed to run on Crossfire since it's the ancient "R300 multiGPU" solution.
 
Didn't ATI try this originally with their super tiling or whatever they called it? They split the screen into a checkboard pattern and had each card render the odd or even tiles.
They just did all the vertex work on both cards, so no triangle could be rendered out of the original order.
Tiling would appear to alleviate any sort of uneven workload, but I'd imagine it's quite a bit more difficult to implement than a simple split frame.
Support for it has to be taken into account during hardware design but it's not difficult.
 
And here I thought someone would paste together a GT300 die pic of a bunch of GT200 units on a big die with an RV770 pasted in the middle.

Back to the topic, it's nearly mid-August. We should be nearing the time where something more concrete about leaked 3dmarks or a slip of something about the design. The long process of building to launch quantity and sampling can't be hermetically sealed from the outside world.
 
Back
Top