Yes, I also thought that it was about perimeter, not area...
AMD RV770 refresh -> RV790
- Thread starter w0mbat
- Start date
Yes, I also thought that it was about perimeter, not area...
DegustatoR
Legend
That's even worse, you'll have to use two different methods of multichip rendering on one board at the same time.
Shintai is only speculating on the pin count dynamics for RV770/RV740, I would not take what he says literally. According to that kind of logic, handheld chips would never be nearly pad limited, yet they can be.
RV740 is a direct competitor to GT216: 512MiB 128-bit 1.8GHz+ GDDR5/16ROPs vs 768MiB 192-bit 1.2GHz GDDR3/12ROPs (and maybe 384MiB for OEMs?) - I don't know the TMU/ALU specs of either however, but ]RV730;RV770] and ]G94;GT200[ seems like a fair bet to me before considering clock speeds; the reason why I say GT200[ instead of G92] is that the absolute maximum I would be willing to seriously consider for GT216 is 32TMUs/160SPs.
Either way RV740 will be a very exciting chip most likely. While it could theoretically be a shrunk RV730 (hey, it could be a shrunk RV250 too!) that makes no practical sense, and I would not exclude the possibility it manages to achieve HD4850-level performance when paired with >2GHz GDDR5. So 2009 will be a very exciting year for 3D - and if I have my way, an even more exciting year for Beyond3D!
EDIT: Since VR-Zone/Expreview linked this post, two quick comments: 1) Once again the TMU/ALU specs are speculation, only take the memory config/ROP specs seriously. 2) 32 TMUs/160 SPs represents 40 SPs (5 "multiprocessors", each 8-wide) per "cluster". As I said in another post, I would *guess* anything from 24 TMUs/72 SPs to 32 TMUs/160 SPs is plausible. 3) What's up with linking to people's brain dumps on forums, even if they include a tiny leak along with all the rest?
RV740 is a direct competitor to GT216: 512MiB 128-bit 1.8GHz+ GDDR5/16ROPs vs 768MiB 192-bit 1.2GHz GDDR3/12ROPs (and maybe 384MiB for OEMs?) - I don't know the TMU/ALU specs of either however, but ]RV730;RV770] and ]G94;GT200[ seems like a fair bet to me before considering clock speeds; the reason why I say GT200[ instead of G92] is that the absolute maximum I would be willing to seriously consider for GT216 is 32TMUs/160SPs.
Either way RV740 will be a very exciting chip most likely. While it could theoretically be a shrunk RV730 (hey, it could be a shrunk RV250 too!) that makes no practical sense, and I would not exclude the possibility it manages to achieve HD4850-level performance when paired with >2GHz GDDR5. So 2009 will be a very exciting year for 3D - and if I have my way, an even more exciting year for Beyond3D!
EDIT: Since VR-Zone/Expreview linked this post, two quick comments: 1) Once again the TMU/ALU specs are speculation, only take the memory config/ROP specs seriously. 2) 32 TMUs/160 SPs represents 40 SPs (5 "multiprocessors", each 8-wide) per "cluster". As I said in another post, I would *guess* anything from 24 TMUs/72 SPs to 32 TMUs/160 SPs is plausible. 3) What's up with linking to people's brain dumps on forums, even if they include a tiny leak along with all the rest?
DegustatoR
Legend
I don't get the [][]][[] brackets system Arun =)
What does it mean?
Why do you think NV won't use GDDR5 in their next middle-end product? (I have my info why this indeed might be the case but i want to hear your version =))
Are you sure? I was under the impression that RV740 would go up against GT214 (as the followup to G94 whereas GT216 is the followup to G96).
No idea how this standard might differ internationally, but [0;1[ for example would be the range including 0 and excluding 1.
Same reason why they didn't go for 55nm immediatly: they likely think the DRAM manufacturers couldn't sustain the necessary capacity for two vendors to switch their entire line-up to it at the same time. In addition to that, they obviously didn't have GDDR5 IP on 65/55, and so releasing their very first product on 40nm with a memory technology they haven't used before is a bit more risky than necessary...
GT21x is GDDR3 all the way except for GT212, while GT3xx-or-whatever-the-hell-it's-called is likely GDDR5 all the way with no exception, IMO.
And G94 is the follow-up to G84, yet I don't think it's in the same level of performance. I'm probably looking at the same data you're looking at, I just think it's indicating price targets more than performance...
In another thread I said this:
And I roughly stick to it, and I think the range of specs I'd find plausible for GT216 is all the way from 3T/72A to 4T/160A. Obviously the former would be a debacle of epic proportions against RV740, while the latter would be very very interesting. We'll see what happens. May I point out that if you take GT200 and scale it down by only a conservative 2x, 32TMUs/160ALUs would take ~30mm² and ~50mm² respectively? If we're talking about a 125-160mm² chip as you would expect with 192-bit GDDR3, this seems very reasonable to me.
Either way I can't imagine a scenario where RV740 isn't a very exciting chip...
And I'd rather not go too OT into NV stuff here....Hah, well that isn't particularly fair to those who do have at least some minimum amount of insider info, isn't it?
What about just starting a new thread for (informed) spec *speculation* in general, for ALL upcoming chips, and leave rumours & architecture into less generic threads? Would that do the job?DegustatoR
Legend
Ah, OK, now i get it =)
Aren't GDDR5 production capacities determined by GPU vendors via demand for GDDR5 at the moment? It's not like GDDR5 is used anywhere besides videocards right now. So if NV want to use GDDR5 in their high volume part they are giving a simple and straight reason for Hynix/Samsung/Quimonda/etc to increase their GDDR5 production capacities.
(My info is more JEDEC-conspiracy-like =))
Well, that's OK if GT212 is the middle end part and everything esle is even lower and top end is filled with GT212x2 AFR board again i guess...
But if RV740 is using 256-bit GDDR5 bus (Fuad seems to be sure that it does) then NV's going to have big trouble matching it with 192-bit GDDR3 bus. And if it's 128-bit GDDR5 (which makes much more sence to me) then they'll probably have a big die size difference again -- especially if 'your' GT216 will use 55nm again.
Jawed
Legend
Honestly it's more of an idle thought, reflecting more on the fact it's extremely difficult to discuss and track what are in effect two entire roadmaps (AMD's and NVidia's for the next 12 months) with a total of 8-10 chips rather than just two individual enthusiast level chips.Hah, well that isn't particularly fair to those who do have at least some minimum amount of insider info, isn't it?
Jawed
And let GDDR3 capacity idle? I doubt they really want that. Furthermore consider that capital expenditures have been dramatically reduced for next year since September...
?????
- GT21x is all 40nm. GT212 is the highest-end chip, GT216 the second-to-lowest-end.
- GT212 won't have a GX2 version, it'll be much too big for that. GT214 could, but won't.
- As for RV740 being 256-bit GDDR5, that's rather ridiculous, isn't it?
DegustatoR
Legend
12T|480A|12R -> 2.8TFlops+ -> GT212/Late June
That doesn't look too big for 40nm, it'll probably end up being smaller than GT200b which will be used for AFR board pretty soon.
Plus i'm thinking it'll be more like 384 ALUs in 12 TPCs for GT212. But that's just me.
Hmm, you might be right. However remember cost/mm² and power/mm² are both going up on 40nm unless you clock things quite a bit lower than they need to be, so there's more to it than it may seem at first. And ofc I hope I'm being conservative in that estimate. Anyway anything more than that and we're going way too OT here so I'll STFU
Fuad is writing a lot of words to fill those pages...
RV740 will be 128-bit with GDDR5 which bandwidth-wise is equal to 256-bit bus at the same frequency.
A11 is normally preproduction, for the samples. A12 is normally first batch production if you're lucky. (The first R6XX generation was very unfortunate)
Recent exceptions: R600 (A13), RV630 (A14/15 IIRC- ouch), RV610 (A15?), RV670 (A11 was good enough)