nvidia "D8E" High End solution, what can we expect in 2008?
- Thread starter 2900guy
- Start date
Megadrive1988
Veteran
D9E is so confusing for most people unless you explain what it means. it's still NV55 as far as I'm counting
G80 was NV50
G92 is NV52 or whatever
D9E could be considered G90 and/or NV55,
the equivalent of the NV47 / G70 refresh of NV40.
Also, I'm not even getting into the GX2 / dual GPU cards, just the upcoming highend refresh GPU, which is not going to be a totally new architecture like G80 was, but the first true refresh.
G80 was NV50
G92 is NV52 or whatever
D9E could be considered G90 and/or NV55,
the equivalent of the NV47 / G70 refresh of NV40.
Also, I'm not even getting into the GX2 / dual GPU cards, just the upcoming highend refresh GPU, which is not going to be a totally new architecture like G80 was, but the first true refresh.
Ok.... Let me see.... NVIDIA is preparing to launch Tri-SLI, G92 doesn`t support it and all G80s GF8800GTS/GTX/Ultra are said to be EOL in January 2008 so it seems that there MUST be next-gen GPU on the way rather than 2xG92.... Maybe TriSLI is for D9E??
Could they introduce TriSLI if the only cards that are supporting it will be EOL??
Could they introduce TriSLI if the only cards that are supporting it will be EOL??
....are you really able to untie that knot? I doubt anyone but NV employees know at this point what is what with all that confusion. Granted there were quite a few changes in the roadmap (no D3D10.2 anymore etc.) and I expect there to be more for the foreseeable future since I personally don't expect M$ to release Vienna in 2009 after all....
In other words in an older roadmap there should have been 2 technology refreshes before D3D11; now there's only one left. The amount of refreshes won't most likely change in the end.
With the way things have been going recently, I wouldn't be surprised if nVidia releases D8E after D9E, with it performing slightly better in SLI favourable situations and having a DX10 feature set rather than DX10.1... Oh, and D8E will be cheaper. At which point they will come out with a new naming strategy for their cards. 
Unknown Soldier
Veteran
News Source: http://www.fudzilla.com/index.php?option=com_content&task=view&id=4763&Itemid=1
-------------------
So, the D8E dual chip in January and then the D9E in March. Interesting. Will wait for March with much interest.
US
Megadrive1988
Veteran
G100 in march? no way, not unless G100 is actually just a refresh of G80, in the same sense that G70 was a refresh of NV40.
G100 is assumed to be the true next new architecture from Nvidia (NV60), and I would expect it until 2009.
so if NV55 / D9E/ GeForce 9 = G100, then okay, but I wasn't thinking that.
NV55 / D9E / GeForce 9 should be G90 (higher-end than G92), but these changes in the way Nvidia names their chips is getting too confusing.
regardless, the GeForce 9 should be a major overhaul / refresh of G80 (and G92). but not a completely new architecture.
Forgive me for a) quoting Fuad and b) being dense and c) redundancy and d) being a jerk, but next-gen high-end in one quarter that's twice as fast as a dual-GPU 8800 series card, or let's say (assuming lower clocks and quad-SLI inefficiencies) ~3x as fast as a single 8800GT? LOL?
Or just another Fuad English slip (by "two [...] dual," he actually means just "two" or "a single-card SLI G92GX2") and his low-detail sources pin G100 at ~2x G80GT--which, while it would still be surprising for a "refresh" part (despite its late arrival) and given G92's size, wouldn't exactly be news for a general "next-gen" high-end part, at least not this far out from G80?
What sort of reasonable architecture can achieve 1.5TFLOPs on a single die?
I seriously doubt that Fudo heard anything about a "G100" codename; it sounds more like his own assumption if someone told him that there's no longer any "G90" in the roadmap.
What used to be known as "G100" was according to my understanding the D3D11 generation chip and for that we still have years to wait for.
NV has obviously changed it's internal codenames to D8x/9x etc.; I know something like "G100" sells better when you rely on sensationalism like Fudzilla, but do we really have to adopt it too?
What used to be known as "G100" was according to my understanding the D3D11 generation chip and for that we still have years to wait for.
NV has obviously changed it's internal codenames to D8x/9x etc.; I know something like "G100" sells better when you rely on sensationalism like Fudzilla, but do we really have to adopt it too?
The problem is neither NVIDIA nor AMD have any good reason to use logical codenames, or even to keep the same ones over time (especially as the engineering teams might use different ones completely). That's why I propose we create our own nomenclature:
N: NVIDIA
A: AMD/ATI
n: nTh Architectural Generation
X: 400-600mm² [Extreme]
E: 275-400mm² [Enthusiast]
P: 175-275mm² [Performance]
M: 125-175mm² [Mainstream]
V: 85-125mm² [Value]
v: 0-85mm² [Value]
In the case a chip is very near the top or the bottom of a range, its target market and pricing should ideally also be considered. As for shrinks, the new process and the revised target segment should be added after the previous codename. So here are a bunch of examples:
R520: A5E
R580: A5E2
RV515: A5V
RV516: A5V-80V
RV530: A5M
RV535: A5M-80M
RV570: A5P
R600: A6X
RV670: A6X-55P
RV630: A6M
RV635: A6M-55M
RV610: A6v
RV620: A6v-55v
G70: N7E
G71: N7E-90P
G72: N7v
G73: N7M
G80: N8X
G84: N8M
G86: N8V
G92: N8X-65E
G98: N8V-65v
As for what's coming out in Q108 and early Q208, I presume that to be N9E/N9P/N9V? As for R680, I'd call that 2*A6X-55P.
N: NVIDIA
A: AMD/ATI
n: nTh Architectural Generation
X: 400-600mm² [Extreme]
E: 275-400mm² [Enthusiast]
P: 175-275mm² [Performance]
M: 125-175mm² [Mainstream]
V: 85-125mm² [Value]
v: 0-85mm² [Value]
In the case a chip is very near the top or the bottom of a range, its target market and pricing should ideally also be considered. As for shrinks, the new process and the revised target segment should be added after the previous codename. So here are a bunch of examples:
R520: A5E
R580: A5E2
RV515: A5V
RV516: A5V-80V
RV530: A5M
RV535: A5M-80M
RV570: A5P
R600: A6X
RV670: A6X-55P
RV630: A6M
RV635: A6M-55M
RV610: A6v
RV620: A6v-55v
G70: N7E
G71: N7E-90P
G72: N7v
G73: N7M
G80: N8X
G84: N8M
G86: N8V
G92: N8X-65E
G98: N8V-65v
As for what's coming out in Q108 and early Q208, I presume that to be N9E/N9P/N9V? As for R680, I'd call that 2*A6X-55P.
I'd love to see such a chip, but I don't think it's doable @ 65nm.
I think 192 SPs @ 2GHz+ with 3 flops/SP (ala G86) is perfectly realistic, but who knows. I'd definitely suspect N9E is basically G80/G92 with architectural tweaks and a higher ALU-TEX ratio. This makes even more sense given that it should be the first chip with FP64 support...
DegustatoR
Legend
256 SPs are doable and the chip should be smaller then G80.
But i doubt the 2GHz part.
I personally think that chip would be pretty big and hard to run at those speeds. AFWIT, AMD/nVIDIA will keep the size lower this generation, in preparation for refresh products with two packages (perhaps on a smaller process in Nvidia's case) like GX2 and R680.
That's why I think they'll be in the realm of ~300-400mm, and at least in ATi's case trying to not to exceed 150W. Nvidia imo has the wiggle room to use 225w on the highend and a bigger die size, as it's replacing the GX2 D8E with imo roughly the same specs quickly, and still has a process shrink down to 45/55nm to look forward to, something ATi won't with R700. I wouldn't be surprised to see a Nvidia refresh/shrink around the time (within a couple months) we see R700 to compete with it.
I agree with Arun's estimation. 192@~2ghz. Something over, but not incredibly exceeding 1Gflop, using as little shaders as possible at the highest comfortable shader clock. Comparable (slightly better) to the D8E in shader power, but higher bandwidth and obviously much more efficient. 2ghz seems to be the current threshold of the comfort zone imho for the architecture on the current process.
Last edited by a moderator:
Unknown Soldier
Veteran
I like your thinking
I think Arun's 2Ghz is shader speed .. not core, which means it's doable.
US