Which RV770 (HD4830/4850/4870)? And which RV870 (mainstream or high-end model)?
AMD: R8xx Speculation
- Thread starter Shtal
- Start date
I'm more than sure Fuad meant 60% more performance if the launch prices are around $200(x850) and $300(x870).
AMD is targeting these price segments and if they would somehow managed to produce a chip 160% faster than RV770, even them would reconsider pricing structure for it ....
Anyway 60% faster than RV770 on average is not bad for first DX11. If it is at most 60% faster than I would be a bit disappointed, but I'm sure that's not the case.
AMD is targeting these price segments and if they would somehow managed to produce a chip 160% faster than RV770, even them would reconsider pricing structure for it ....
Anyway 60% faster than RV770 on average is not bad for first DX11. If it is at most 60% faster than I would be a bit disappointed, but I'm sure that's not the case.
Well, I see that a majority of arguments do speak against my point of launching a salvage part and the second-fastest configuration simultaneously, saving the surprise box, i.e. the fastest SKU for later. I am still not a hundred percent convinced, but I see your valid points.
auntie edit says: Just wanted not leave the previous string of discussion without any remark about me possibly being wrong
auntie edit says: Just wanted not leave the previous string of discussion without any remark about me possibly being wrong
Which could theoretically be amended by letting the salvage parts use full mem controllers but populate the SKUs with GDDR3 only.
Last edited by a moderator:
Jawed
Legend
If anything I imagine Juniper has been in production longer than Cypress - it's smaller/easier (i.e. more like RV740) and AMD needs more of them for the launch, since it seems AMD's going to launch the two together. That would add another factor into the volume shortfall facing HD5830.
It seems an L2 and RBE are both associated with a sole memory controller.
You're talking about a situation where all the MCs work but some L2s or RBEs don't? Taking a single MC-L2-RBE unit, the L2 might be working but the RBE not, or the RBE is working but the L2 not. I suppose it would be possible to get a driver to support such asymmetric configurations (since L2 is for textures and so doesn't have a connection with the local RBE). But it's pretty strange, as it would require two different kinds of tilings, one for textures and one for render targets. Maybe the tilings are easy anyway, due to the overall architecture - I don't know.
Jawed
From what I hear, maybe Cypresses have been growing faster this summer than other evergreen plants.
WRT to salvage parts: I was implying no defects on the L2/RBEs at all - given how hard it must be to find defective parts, I'd have guessed the less defects are required, the better.
WRT to salvage parts: I was implying no defects on the L2/RBEs at all - given how hard it must be to find defective parts, I'd have guessed the less defects are required, the better.
They do that? so far we've seen numbers that do not align with "poor and weak" same goes for RV770.
DUH It's reverse-reverse psychology. You think I'm strong, while I'm portraying weakness only to be smitten by my über-strength!
DUH It's reverse-reverse psychology. You think I'm strong, while I'm portraying weakness only to be smitten by my über-strength!
feigning weakness yes.. but so far we've seen numbers that should instill fear in everyones heart. exactly how many low-end videocards score P4k+ in Vantage? We're still missing Hemlock figures as well.
Some thoughts...
a. RV770 : 55nm ---> 256 mm²
b. "RV770b" : 40nm ---> ~145mm²
c. RV870 : 40nm ---> ~320mm² (Chiphell)
d. So RV870 = ~2.2 x "RV770b"
e. RV770 = 956 million transistors, so RV870 = ~2100 million transistors
f. RV770: (TMUs+ALUs+RBE) = ~60% die, (UVD/MC/PICe/CrossFire/Display controllers/Xilleon/Audio/...) = ~40% die
g. "RV770 x 2" = 956 + ~60% = ~1530 million transistors ---> 1600 SP, 80 TMUs, 32 "ROPs"
h. ~2100 million transistors - 1530 = 570 million transistors.
i. 570 million transistors ---> (DirectX 11 implementation) or (800 SP+40 TMUs+16 "ROPs")
j. So RV870 can have (1600 SP+80 TMUs+32 "ROPs"+570M transistors for DX11) or (2400 SP+120 TMUs+32 "ROPs")
k. Maybe I think too much.
a. RV770 : 55nm ---> 256 mm²
b. "RV770b" : 40nm ---> ~145mm²
c. RV870 : 40nm ---> ~320mm² (Chiphell)
d. So RV870 = ~2.2 x "RV770b"
e. RV770 = 956 million transistors, so RV870 = ~2100 million transistors
f. RV770: (TMUs+ALUs+RBE) = ~60% die, (UVD/MC/PICe/CrossFire/Display controllers/Xilleon/Audio/...) = ~40% die
g. "RV770 x 2" = 956 + ~60% = ~1530 million transistors ---> 1600 SP, 80 TMUs, 32 "ROPs"
h. ~2100 million transistors - 1530 = 570 million transistors.
i. 570 million transistors ---> (DirectX 11 implementation) or (800 SP+40 TMUs+16 "ROPs")
j. So RV870 can have (1600 SP+80 TMUs+32 "ROPs"+570M transistors for DX11) or (2400 SP+120 TMUs+32 "ROPs")
k. Maybe I think too much.
Last edited by a moderator:
a. RV770 : 55nm ---> 256 mm²
b. "RV770b" : 40nm ---> ~145mm²
c. RV870 : 40nm ---> ~320mm² (Chiphell)
d. So RV870 = ~2.2 x "RV770b"
e. RV770 = 956 million transistors, so RV870 = ~2100 million transistors
f. RV770: (TMUs+ALUs+RBE) = ~60% die, (UVD/MC/PICe/CrossFire/Display controllers/Xilleon/Audio/...) = ~40% die
g. "RV770 x 2" = 956 + ~60% = ~1530 million transistors ---> 1600 SP, 80 TMUs, 32 "ROPs"
h. ~2100 million transistors - 1530 = 570 million transistors.
i. 570 million transistors ---> (DirectX 11 implementation) or (800 SP+40 TMUs+16 "ROPs")
j. So RV870 can have (1600 SP+80 TMUs+32 "ROPs"+570M transistors for DX11) or (2400 SP+120 TMUs+32 "ROPs")
k. Maybe I think too much.
At least one thing to change that formula a bit is that I think someone mentioned that for example mem controllers don't scale as well as many other parts of the core do when you move to smaller process, but would it have significant effect, I don't know
Only if things worked like that and ATI/Nvidia's R&D costs would shrink so much.a. RV770 : 55nm ---> 256 mm²
b. "RV770b" : 40nm ---> ~145mm²
c. RV870 : 40nm ---> ~320mm² (Chiphell)
d. So RV870 = ~2.2 x "RV770b"
e. RV770 = 956 million transistors, so RV870 = ~2100 million transistors
f. RV770: (TMUs+ALUs+RBE) = ~60% die, (UVD/MC/PICe/CrossFire/Display controllers/Xilleon/Audio/...) = ~40% die
g. "RV770 x 2" = 956 + ~60% = ~1530 million transistors ---> 1600 SP, 80 TMUs, 32 "ROPs"
h. ~2100 million transistors - 1530 = 570 million transistors.
i. 570 million transistors ---> (DirectX 11 implementation) or (800 SP+40 TMUs+16 "ROPs")
j. So RV870 can have (1600 SP+80 TMUs+32 "ROPs"+570M transistors for DX11) or (2400 SP+120 TMUs+32 "ROPs")
k. Maybe I think too much.
I think the real wrench in such calculations is the new scheduler.
Only if things worked like that and ATI/Nvidia's R&D costs would shrink so much.
But it would be nice for all of us layman speculators
Yep the old school way like on RV200-RV570 (1650XT) kind of style. And nope ATi forbid it and probablynever will. Maybe some HD4750 w/ 1gb gddr3/4
Retail HD4770 definitely crappy cooling but the original one fit's it better than on these Junipers
please... please.. for the love of god, delete that! Shirley, You can do better with MSPaint!
They were create here at B3D by Wirmish and tell me, You've read the thread @TH?
Similar threads
