What about it? The likelihood of it coming before GT300 is...ugh...very small, to say the least.
AMD: R8xx Speculation
- Thread starter Shtal
- Start date
What about it? The likelihood of it coming before GT300 is...ugh...very small, to say the least.
True. So it makes more sense to fine-tune a chip's size to fit a 256-bit bus on it.
Which might mean much more work later on to shrink it or maybe even design a new chip for a shrink.
Another ~250mm2 chip for a 256bit bus would be interesting. ~1.6-1.7b trannies, straight shrink to 32nm would be ~165mm2, adding on some fat since shrinks are not linear, looking around 180-190mm2. Just enough for a 256bit bus? Replaces RV840/RV870Pro next summer/fall?
Jawed
Legend
Which is why in my view this isn't the biggest GPU of this family.
According to B3D, R420 was 281 and R520 was 288mm²:
http://www.beyond3d.com/resources/chip/63
http://www.beyond3d.com/resources/chip/95
It's the GPU that you'd expect to pay <=$100 this autumn. ~50% more performance than HD4770 with D3D11, ~6 months after HD4770 launched - or maybe ~10 months or more after HD4770 was supposed to launch.
A you suggesting that ATI will opportunistically abandon the large chip it was planning for release this year solely because of a rumour that NVidia's late?
Or are you suggesting that ATI never intended to have a part with significantly more performance than RV790 until 2010Q2?
More pertinently, why is RV740 considerably larger than it need be for a 128-bit GPU?
Maybe they're targetting performance, not die size?
More fundamentally, can they design for die size? RV770 was designed to be 256-bit GDDR5 but was apparently expected to be bigger than it turned out (though CrossFireX Sideport might have come along and buggered that up since it's pretty clear it could be much smaller without that and with only the original 8 clusters).
The design work for RV770 started way before 55nm process was available - so the libraries were an unknown factor until some later time. It seems to me that if you're designing for "being early" on a major node you can't design for specific die sizes. Libraries plus the properties of a node must add a noticeable error margin on die-sizing.
This makes it more likely, in my view, that this Evergreen chip is 128-bit - it's cutting it really fine to fit a 256-bit interface on there purely in terms of the perimeter required for 256-bit GDDR5, according to the RV770 die shot.
Jawed
192-bit makes sense to me. I'm inclined to believe the 25:1 Alu:Tex and 1200:48 specs just because they have been thrown around so much. If they use 1.2Ghz GDDR5, they would have as much bandwidth as the 4870. So you would end up with:
Performance @ 900MHz vs HD4890:
24 ROPs +58%
48 TMUs +27%
1200 ALUs +58%
768 MB GDDR5 -25%
115.2 Gbps bandwidth -8%
I'd estimate performance vs HD4890 at +25 to 40%. Could comfortably sell for $199, but as it will be DX11 and faster than the HD4890, probably at least start at $249 (possibly even $299 with no real competition). Maybe a 725MHz variant starting at $199. And when the competition finally arrives, reducing prices should be fairly painless.
Performance @ 900MHz vs HD4890:
24 ROPs +58%
48 TMUs +27%
1200 ALUs +58%
768 MB GDDR5 -25%
115.2 Gbps bandwidth -8%
I'd estimate performance vs HD4890 at +25 to 40%. Could comfortably sell for $199, but as it will be DX11 and faster than the HD4890, probably at least start at $249 (possibly even $299 with no real competition). Maybe a 725MHz variant starting at $199. And when the competition finally arrives, reducing prices should be fairly painless.
Oh, so I must've googled it wrong
Anyway, RV740 is a good example: it is a successor of RV730, which is a bit bigger, and competes with G94 and G92. RV670 was also rather small, although there isn't a chip I could directly compare it with.You probably won't argue that G73 was a direct successor to NV43. According to B3D, NV43 is 150 mm² and G73 is 127 mm².
I can't rule out neither.
I wouldn't make a big deal out of it, there were bigger 128bit GPUs in the past.
Interesting... so you're telling me that RV670 being the perfect size for a 256bit bus was just ATI getting very lucky?
=>ninelven: 192-bit bus would certainly look nice on it. People complain that 512 MB of memory is not enough, so here you go, 768 megs. Also, since this would be the slower part as Jawed is suggesting, they could make a cheaper GDDR3 or GDDR4 equipped variant. But somehow I don't think ATI will do this...
R520 used a pretty mature manufacturing process cause RV410 (X700Pro) used it as small chip competing NV43 (6600GT) whole year before R520 came out to market. And then there was R430, not so successful refreshment of R423, but higher-end chip comparing to RV410 that was a proving ground to build R520 around it. And afaik they want to deliver R600 as R500 architecture with 9.0c GPGPU compatibility. So that's why R520 carried that name cause it was a lineage delivered with first R100 with great trilinear filtering. Something that lacks in R600 design beyond poor AF comparing to all G80 based products.
So that's why i mention above these RV870 something as R520 which just add 9.0c compatibility to R420 based product line whole 18 month later.
You missed part of thread. Nio, i'm suggesting that it seems that ATi will have something like RV670 (in size) with 200USD MSRP if this is a real presentation. And that real RV870 wannabe will come in 6-9 month time (i repeat myself) around May 2010. And AMD will have great sales with 99,9999% of people just wanting new tech and it will offer it cause DX11 games (patches for todays) will certainly run much faster on new RV870 than on RV790 even if they had the same amount of SP (not 50% more) and run at the same frequency.
After all great thing could be that ATI could finally have a chip that wouldn't lack of power inside it's core considering today available PWM management on 200$+ cards like HD485 (HD4870 nowadays).
And it would have some 100W envelope instead 160W (RV770)/190W (RV790) and delivering double performance with DX11 with same scene detail
I like that but why you all put only 768MB on that card. If it's upper mainstream it could have 1536MB just like nV gtx275 successor and what's with all that memory bandwidth limitation ... Qimonda had their 1250MHz GDDR5 chips for half year Hynix has listed 1.125MHz parts (even they announced in Nov2007 that they produced 1250MHz 1Gbit parts) and Samsung also has 1125MHz and until Sep2009 who knows Samsung announced 1750MHz 50nm chips in Feb this year. So probably they all have more advanced chips if there will be a demand in next year
GDDR5 per chip 32b chip 192b bus 256b bus
MHz Gb/s GB/s GB/s GB/s
800 3.2 12.8 77 102
900 3.6 14.4 86 115
1000 4.0 16.0 96 128
1125 4.5 18.0 108 144
1250 5.0 20.0 120 160
1500 6.0 24.0 144 192
1750 7.0 28.0 168 224
Last edited by a moderator:
I think you're wrong. To my knowledge, RV410 and R430 both used the 110nm process. R520 was on 90nm. Of course, ATI wouldn't risk it if they hadn't already tried the 90nm with Xenos, but still it was the first 90nm desktop GPU.
R520 didn't just add DX9.0c compatibility, there were some architectural changes, that later made it possible to multiple the shader units (for R580) or run GPGPU tasks.
It seems I'm
but i still believe that was 110nm for R520 also, and only R580 used 90nm. Well i must remember that . Well then it was much higher packed chip considering 90nm is full node down to R420 130nm, and twice as much space.And about that 128b bandwidth. They don't use wider pcb just cause they can because it also cheaper than to wait and depend on only one source. Qimonda was in financial troubles so they maybe get a better deal with them last year, but just to skip to 128b starved node and to wait for Samsung to put a price on theirs high end chips would be a lunacy. And most today's gddr5 could sufficiently get the job done at 192bit bus with just as same bw as RV770 @1250 and i don't believe new RV870 would be a bandwidth starved with just double processing power, and probably with all that threading that doesn't need to be cached now, just put into next SP if that's the case with tessellation.
Jawed
Legend
For what it's worth, I agree that 40nm can make successor chips smaller - I was just querying the specific high-end chips you were referring to.Oh, so I must've googled it wrong
You probably won't argue that G73 was a direct successor to NV43. According to B3D, NV43 is 150 mm² and G73 is 127 mm².
You said:
and my remaining issue with this is that ~100mm reduction in die size is basically deleting all the extra die area that it's possible to use with the new process, area that would have been used in order to increase per chip performance.
Extremely weak I'd say. You're arguing that top end ATI chips won't have a significant performance increment after ~15 months.
You think G92 v G94 syndrome is the prime motivator here, while I think it's unlikely the chip we've seen is 256-bit and even if it were, there's a possibility that the new high-end chip has 32 RBEs - though I don't rate that possiblity highly.
If the chip we've seen is a radical departure in architecture and is the holy grail of multi-chip sharing, well that screws up everything
Yet this sector of the market is more sensitive to price and so die size should be a huge factor, supposedly.
What makes you think it's the perfect size? It's bigger than this new chip, which you're arguing has a 256-bit bus (I think it's just about possible - but I honestly don't know how close). Not only that but GDDR5 takes more space for 256-bit than RV670's GDDR3.
The bottom line is how limited by bandwidth Evergreen GPUs are. If GDDR5 isn't scaling very fast then that puts a cap on the non-ALU performance improvement over RV790. Then it's a matter of the degree to which RBE count is constrained - are 32 RBEs likely and could they bring a significant performance improvement even within the 256-bit GDDR5 bandwidth constraint?
Does Evergreen have RBEs?
Jawed
It seems I'm
Umm, no? Lecithin brotha', do you take it?
I could be wrong but I keep hearing about large 1100 GGDR5 orders lately....
Don't know but by the time I read about a hypoethetical 192bit bus and 24 ROPs, something closer to 200 than 240 ALUs jumps into my mind and yes that's just me.
FYI:
http://vr-zone.com/forums/441732/evergreen-cypress-and-juniper-ati-code-names-explained.html
It seems that those ~180mm^2 dice were Cypress (formerly known to us as RV870)
http://vr-zone.com/forums/441732/evergreen-cypress-and-juniper-ati-code-names-explained.html
It seems that those ~180mm^2 dice were Cypress (formerly known to us as RV870)
Jawed
Legend
That's an article from semi-accurate:
http://www.semiaccurate.com/2009/06/09/ati-evergreen-code-names-explained/
Also, the original video I linked, as posted by amdunprocessed, showed "EG BROADWAY":
http://forum.beyond3d.com/showpost.php?p=1298835&postcount=723
http://forum.beyond3d.com/showpost.php?p=1298862&postcount=727
So Broadway is a codename that needs explaining too.
Jawed
Jawed
Legend
So that one should be "R800"? Especialy since now is bit unlikely that "RV810" will take part in these discussions
Maybe due to the density of RV670?
They certainly increased density quite a bit afterwards with RV770.
I should hope it was not R800, that demo wasn't exactly mind blowing and it was only getting what some would call "mediocre" performance of about 20-30FPS.
Similar threads
