AMD: R7xx Speculation

Status
Not open for further replies.
  • Chip was pad limited in the beginning***, so the last couple of SIMDs are value adds that weren't originally planned for. Explains the first point a little bit.

That's really puzzling. The way thing normally go, final area is always larger than initially estimated. I've never seen it otherwise...

This really points to some serious design methodology changes along the way.
 
FP16 filtering is half speed, and the samplers are limited by available interpolators (only 32 texcoords/clk) when processing INT8 at full speed
Now, I know what I want to see *fixed* in the next full/half node refresh of the architecture! :D
 
Wait, what is the "96 Shader" chip Orton talked about awhile ago? I'm confused. You mean the 480 SP RV770 that turned out to be a ruse?

That's really puzzling. The way thing normally go, final area is always larger than initially estimated. I've never seen it otherwise...

This really points to some serious design methodology changes along the way.

Starting point was the cost target. The initial specification was 150% over R600, however the engineering focus was area optimization, so once they got underway the efficiency gains mean we could be 8 SIMD's (200% R600); a little while later, due to further gains, GDDR5 and a couple of other things, we were pad limited at 8 again, so the additional 2 SIMD's were lobbed in. And I'm not complaining about that! :)
 
FYI, the 4870 review thread is here.

Anyway, if this thread stays open until Rys publishes B3D's piece, I'd like to know what's up with RV770's idle draw. People here who seem to be in the know have said it's the drivers, but I haven't seen it mentioned specifically in any p/review, not even the new ones, so has AMD officially briefed ppl about it?
 
FYI, the 4870 review thread is here.

Anyway, if this thread stays open until Rys publishes B3D's piece, I'd like to know what's up with RV770's idle draw. People here who seem to be in the know have said it's the drivers, but I haven't seen it mentioned specifically in any p/review, not even the new ones, so has AMD officially briefed ppl about it?

I think people are basing it on the idle speed of the rv770, its idling at 500mhz instead of the 300 that the rv670 idles at. I don't know that it's a bug.
 
Starting point was the cost target. The initial specification was 150% over R600, however the engineering focus was area optimization, so once they got underway the efficiency gains mean we could be 8 SIMD's (200% R600); a little while later, due to further gains, GDDR5 and a couple of other things, we were pad limited at 8 again, so the additional 2 SIMD's were lobbed in. And I'm not complaining about that! :)

Could you expand a bit on what you mean by "area optimization"? Was the optimization at an architectural level (as in changing the function and structure of units) or at a synthesis level (making the synthesizer choose smaller cells, improve area utilization and routing, using custom standard cells (FAST14????)).

Also, can you tell us if parts of the chip were built full-custom? It would seem like the shader units would be the perfect target for this. I had a theory that in R600/RV670, the shaders were synthesized/semi-custom while in RV770 the shaders are full-custom. This seems somewhat plausible due to the vast improvement in area that full-custom offers. Is there any truth to this?
 
Status
Not open for further replies.
Back
Top