Yes.
The only concerning thing for Nvidia is they're again competing with ATI with a disadvantage. RV740 is what .. half the size of G92b. Also wonder if the new GTS240 is dual-slot, 8 layer PCB design?
AMD RV770 refresh -> RV790
- Thread starter w0mbat
- Start date
The only concerning thing for Nvidia is they're again competing with ATI with a disadvantage. RV740 is what .. half the size of G92b. Also wonder if the new GTS240 is dual-slot, 8 layer PCB design?
DegustatoR
Legend
Considering that they're on two completely different TSMC nodes that's not necessarily a disadvantage. In the long run - yes, during 2009 - unknown.
Yes, got to factor that in too. Did Nvidia indicate if the 'stock piles' of inventory is G9x or GT200?
they said 65nm parts.
Cookie Monster
Newcomer
TBH, i prefer dual slot coolers to single slot coolers nowadays. Much more efficient and with onboard hardawre improving alot on motherboards, there's not much need for PCI/PCI-e x1 slots except for those few professionals that require raid cards/soundcards etc. (Im just wondering when nVIDIA will move onto digital PWM with their PCBs..)
So basically, RV740 will have performance around the HD4830? What are its rumoured specs? The GTS240 is basically a 9800GTX, so it packs quite abit of punch compared to the 9800GT.
740XT should lead 4830 by a discernible margin. 
;P
640/32 is basically the same, just that this time there's an unignorable 20% clock boost. Should just about translate to improvements in the teen percentages.
An overclocked RV740 by say, 10-15%, should beat the GTS 240 all-round. But I would be eyeing the leakage concerns first.
The new product matrix might be a little bit scary...
RV790 XT vs GTX 285, probably loses but does a coup anyway.
RV790 PR vs GTS 260 216 (if it exists and uses GDDR5 still)
RV770 XT is all alone, EOL soon
RV770 PR goes down in pricing together with the GTS 250, perhaps even more.
RV740 XT vs GTS 240, fine line to thread here since the 250 is only slighty more expensive.
RV740 PR vs 9600GT, more or less a win at lower resolutions. Not sure about the bandwidth tax. Notebook new favourite?
;P640/32 is basically the same, just that this time there's an unignorable 20% clock boost. Should just about translate to improvements in the teen percentages.
An overclocked RV740 by say, 10-15%, should beat the GTS 240 all-round. But I would be eyeing the leakage concerns first.
The new product matrix might be a little bit scary...
RV790 XT vs GTX 285, probably loses but does a coup anyway.
RV790 PR vs GTS 260 216 (if it exists and uses GDDR5 still)
RV770 XT is all alone, EOL soon
RV770 PR goes down in pricing together with the GTS 250, perhaps even more.
RV740 XT vs GTS 240, fine line to thread here since the 250 is only slighty more expensive.
RV740 PR vs 9600GT, more or less a win at lower resolutions. Not sure about the bandwidth tax. Notebook new favourite?
Last edited by a moderator:
I wonder whether or not all the rumors floating around about green editions of 9800 GT/GTS240 and 9600 GT (with no new name ) have something to do with 40nm-problems.
Anyone else got the feeling, Nvidia will be trying to emphasize lower power consumption (albeit also lower performance FWIW) as long as they do not have their own 40nm-GPUs ready? I mean, with a new name and DX11 not immediately coming out, plus no need for additional power connectors - maybe opposed to RV740, !!!dunno, rumors!!! - they surely could score some OEM design wins for back to school season, right?
For the end user though it feels kinda lame.
) have something to do with 40nm-problems.Anyone else got the feeling, Nvidia will be trying to emphasize lower power consumption (albeit also lower performance FWIW) as long as they do not have their own 40nm-GPUs ready? I mean, with a new name and DX11 not immediately coming out, plus no need for additional power connectors - maybe opposed to RV740, !!!dunno, rumors!!! - they surely could score some OEM design wins for back to school season, right?
For the end user though it feels kinda lame.
Pantagruel's Friend
Newcomer
740XT should lead 4830 by a discernible margin.
640/32 is basically the same, just that this time there's an unignorable 20% clock boost. Should just about translate to improvements in the teen percentages.
However, it seems like you'll have half as many render backends. I think it will easily counter the clock advantage...
well, this downclocking does sound kinda lame. heck. I've just rebranded my HD4850 to a HD4870 (red) Green Edition! It's the same VPU and uses less power!
Combining it with the 40nm rumours it does sound like both parties have problems getting their power budget under control. the green edition would seem that a underclocked 55nm GPU will use less power than the original clocked 40nm GPU.
Combining it with the 40nm rumours it does sound like both parties have problems getting their power budget under control. the green edition would seem that a underclocked 55nm GPU will use less power than the original clocked 40nm GPU.
According to VRZ's info RBEs remain at 16.
But really, even with 8 RBEs RV730 did pretty decent against the RV670s. Yes, it could have been even better, but for that much ALU power and bandwidth it was quite an accomplishment already.
Pantagruel's Friend
Newcomer
Yes, I've seen both versions many times. Two things that may point towards the 8RBE version:
- Wouldn't the 16RBEs automatically mean a 256-bit bus? I'm not sure how much flexibility is there.
- With 640 ALUs, 32 tex units, 32 interpolators (?) and 16RBEs, this chip is hardly less transistors than the RV770 - AMD surely don't aim for 800-850M trs?
Then, the 8 RV700 RBEs equal the 16RBEs of the RV600 line in several aspects - no halving of resources here. While the 640/32/32/8 configuration effectively doubles the ALU:RBE ratio - I'm pretty sure the backends will seriously bottleneck the performance.
16 RBEs would mean 64 ROPs.
Jawed
Legend
I was just looking at:
http://www.driverheaven.net/reviews.php?reviewid=718&pageid=1
where HD4670 is outclassed by 9600GT. What's the die size of 9600GT with 55nm G94b?
http://translate.google.com/transla...&hl=de&ie=UTF-8&oe=UTF-8&prev=/language_tools
On page 3 currently there's a news item about the Sparkle 9600GT: 197mm2.
Adding just bandwidth to HD4670, with no increase in core clock or RBE capability, it's not going to compete with 9800GT/9800GTX in absolute performance, it will only be competitive with 9600GT. Of course it'll be teeny, but it seems to me AMD's going to have a vast gap in its contemporary line-up without the combination of both RBEs and bandwidth.
AMD really needs an RV7xx GPU at ~200mm2 with a 256-bit bus. This is the argument for a 40nm RV790...
Jawed
I don't think it would be too much work to change the quad-RBEs to attach to 32bit channels instead of 64bit channels. The other question of course though is if it would make sense before the RBEs are completely bandwidth limited - sure they've got caches but at some point that won't really help (of course, the rv770 paired with gddr3 has a similar "high RBE capability to mem bandwidth ratio").
Why not? Looks to me like such a chip could still fit into a similar die size as rv730.
The question though is if the bottleneck isn't rather the memory bandwidth anyway.
Pantagruel's Friend
Newcomer
Ummm, OK, I was trying to stick to ATI terminology. You're right that the groups of four are called render backends - but what are the things that nVidia calls ROPs called at ATI/AMD?
Pantagruel's Friend
Newcomer
OK - I'm not arguing here.
Yes, if that die size is the target. I heard rumours about 100mm2, which would mean around 700M trs.
I have a kind of answer to that - I've done a test with a 4870 clocked at 625/900 and compared it with a 4850. You can see it here - it's not in English, but the graphs are international. Focus on the AA cases as the 4870 kicks the CPU's ass badly :smile:
It seems to me that the memory indeed bottlenecks the 4850, but the 80% extra BW rarely brings more than 10% extra performance. Combined with the observation that a similarly clocked 4830 and 4850 perform almost the same in many cases, I dare say that the backend is bottlenecking, too.
Ummm, OK, I was trying to stick to ATI terminology. You're right that the groups of four are called render backends - but what are the things that nVidia calls ROPs called at ATI/AMD?
ROPs?
Ok, sounds a bit tiny for middle-of-the-pack gpu though (even the uninspiring rv630 and rv635 were quite a bit larger)
Not really arguing with that, but don't forget there are other potential issues why 4830 and 4850 could perform similar, one being that it could be setup limited.
Though I agree that 8 RBEs would quite likely limit performance at least to some degree.
Seems to me the rumors aren't quite converging yet - if it should achieve HD4830-like performance (though model number doesn't exactly suggest this) then 640SP, 32TMU, 16 RBEs (and 128bit, 900Mhz GDDR5) sounds logical. But no way that's going to fit into a 100mm² die.
Jawed
Legend
RV770's 10 clusters are ~40% of the die - let's call it 390M transistors. 8 of those for RV740 would be 312M.
Stretching, somewhat, based on the layout of an RV770 cluster, an RV730 cluster (40 ALU lanes, 4 TUs) is about 25M transistors (while an RV770 cluster is about 39M).
So RV730's 8 clusters consume 200M transistors out of a total of 514M. That leaves 314M for RBEs, L2, MCs and other gubbins.
If we take the 314M for gubbins from RV730 and add it to the 312M for clusters we get 626M.
What's worth noting is the densities at 55nm, in millions of transistors per mm2:
So a 100mm2 RV740 has space for about 600M transistors, which means RV740 has the same RBE count as RV730.
QED
So, is RV740 really that small?...
Jawed
Stretching, somewhat, based on the layout of an RV770 cluster, an RV730 cluster (40 ALU lanes, 4 TUs) is about 25M transistors (while an RV770 cluster is about 39M).
So RV730's 8 clusters consume 200M transistors out of a total of 514M. That leaves 314M for RBEs, L2, MCs and other gubbins.
If we take the 314M for gubbins from RV730 and add it to the 312M for clusters we get 626M.
What's worth noting is the densities at 55nm, in millions of transistors per mm2:
- RV770 256mm2 956M - 3.7M
- RV730 146mm2 514M - 3.5M
- RV710 73mm2 242M - 3.3M
So a 100mm2 RV740 has space for about 600M transistors, which means RV740 has the same RBE count as RV730.
QED
So, is RV740 really that small?...
Jawed