The Official RV630/RV610 Rumours & Speculation Thread

Status
Not open for further replies.
Well most of the reactions to today's 8600 GTS reviews would suggest that bandwidth efficiency hasn't improved to the extent that gamers on a mid-range budget would have liked.

And though I don't expect it to happen, I can also see where AMD could have scored big in retail mindshare with a 65nm 256-bit offering in the $250 price range.

I not except for may launch any 256bit card in 200-230$ range, my quess is late august, BTS.
 
So, in terms of unit counts (die area) G84 is 1/4 of G80 (SPs and TMUs), but with 1/2 the texture addressing and 1/3 ROP/MCs. What are the chances that RV630 is 1/4 of R600? I'd say slim, it's more likely that RV610 is 1/4.

It seems likely that RV630 is half of R600 in terms of unit counts.

I think that might explain the rumoured relatively high power consumption of HD2600XT, ~50W more than 8800GTS.

Jawed
 
Can we try to compare die sizes of G84 and RV630?

g84-chip2.jpg




r6xxgpusxb4.jpg


RV630 is to the most right of the chips.

Maybe we can tell by die size which chip has a bigger engine under the hood.
 
Not bad...but the Rv630 is less than a half R600. It has only a fourth of the memoryinterface, and only a fourth of its ROPs.

~50W more than 8800GTS.

It is never ~50W...

Rv630:

32 5D US
16 BiTMUs (maybe only 8)
4 ROPs
~700Mhz Coreclock (I expect less)

mfg Nakai
 
Not bad...but the Rv630 is less than a half R600. It has only a fourth of the memoryinterface, and only a fourth of its ROPs.



It is never ~50W...

Rv630:

32 5D US
16 BiTMUs (maybe only 8)
4 ROPs
~700Mhz Coreclock (I expect less)

mfg Nakai

And who say's this?
 
Correct me if I'm wrong. But looking at the surface area of G84 and RV630, it seems RV630 is ever so slightly bigger. Then you must take into account of RV630 being 65nm. Unless I'm off, their has to be something packed in that die.
 
Of course, if R(v)6xx architecture lacks multiple clock domains and has similar ALU structure, it would need ~2x ALUs to achieve the same shader performance, assuming perfect frequency scaling.
 
Concensus guesses put RV630 at 160mm2, as far as I can tell. That's 65nm.

G84 is 169mm2 on 80nm.

If RV630 was implemented in 80nm, pure scaling (which to be honest I think is dodgy) would make it 240mm2. That's about 57% of the size of R600's 423mm2 (is that the right figure?). And, way bigger than G84 (some 42%).

But RV630's memory bus is "known" to be 1/4 of R600's, 128 v 512-bit. So that's a reduction in die space given to MCs. Which means either more space for other gubbins, or that 65nm scaling is considerably lower than the ratio of the processes (65nm:80nm).

Jawed
 
If the ROPs are connected like the G80´s, with increasing number of ROPs the Memoryinterface get bigger too. And if the Rv630 and the R600 uses a Ringbus-Controller, the Rv630 is only a fourth R600, but only in the view of memoryinterface and the number of ROPs.

mfg Nakai
 
What's the smallest that RV610 could be?:
  • 16 ALUs (vec5 MAD), 4:1 ratio ALU:TMU
  • 4 TMUs
  • 4 ROPs
  • 2 MCs (32-bit) = 64-bit memory
After some measuring based on RV610 and RV630 pictures (various board and die shots), I'm pretty sure that RV610's die is 50% of the size of RV630's.

So, ahem, based on the smallest RV610, RV630 would seem to be, at least:
  • 32 ALUs (vec5 MAD), 4:1 ratio ALU:TMU
  • 8 TMUs
  • 8 ROPs
  • 4 MCs (32-bit) = 128-bit memory
Jawed
 
Do we have die size data for any 65nm GPU that has a functionally identical counterpart at an older process? This would give us a percentage size reduction baseline.

EDIT: Well, the closest I was able to find was a quip at Fudzy about G78, which he says is a ~60mm^2 65nm G72, which is of no huge help since G72 is a 90 nm chip. Still, if we assume it's exactly 60 mm^2, it would give us a 22% die size reduction Vs 28% theoretical one.
 
Last edited by a moderator:
It's not exactly normal to see a dual slot cooler come stock on a mainstream card. That card has to be packing something under that hood.
 
Status
Not open for further replies.
Back
Top