Printable View
We could also take into consideration the possibility of an asymetric 2-chip GPU : 700mil transistors in total, 1.5Ghz in total, two different memory controllers with different memory configurations, hybrid water/air cooler (different for the two chips) ... SLI technology at it's best :). And that is doable on 90nm.
two chip GPU, it'll make the board more expensive but oh well.Quote:
Originally Posted by Micutzu
And probably not sporting unified shaders
Quote:
* Close to 700M transistors (G71 - 278M / G70 - 302M)
As other have said
I doubt this is possible
Would be sweet if im wrong:woot:
Agreed with nn step, there is something that's not included in the specs somewhere. Highly suspect this is a 2 chip solution, but why the odd amount of memory then? 384/2...
Double the clockspeed and you "could" have double the shader power without adding more shaders, so 2x from a chip with twice the clockspeed and twice the transistors (roughly) again seems like something is wrong. However, specs that are out and out wrong, rarely get pulled...
A 2 chip solution however, has the inherent issues of requiring SLI to function. SLI works well, however, not perfectly. It does seem to be a dual chip solution somehow, as the #'s are highly suspect to be about twice as high as 90nm or 80nm would allow.
We'll see.
im willing to bet it has 2 chips, one main with 256-bit memory interface and 512mb GDDR4Quote:
* 384-bit memory interface (256-bit+128-bit)
* 768MB memory size (512MB+256MB)
* Close to 700M transistors (G71 - 278M / G70 - 302M)
then the other with 128-bit interface and 256mb GDDR4
then for the less expensive card, the other chip has its memory interface cut in half to 64-bitQuote:
GeForce 8800GTX : 7 TCPs chip, 384-bit memory interface
GeForce 8800GT : 6 TCPs chip, 320-bit memory interface
Main processing core plus secondary post processing unit?
any release dates?
that would explain the new AA methodQuote:
Originally Posted by onewingedangel
could people pls drop the bit about 1.5Ghz core not possible etc.
we dont even know how accurate all these specs are.
And they dont say the core will be clocked at 1.5Ghz but will be scallable upto with this architekture wich doesnt mean that wel actualy see it happen.
And btw most likely there will also be a 65nm refresh of it at one point, so if they go high core clock near 1.5Ghz it will likely be with the 65nm refresh and not with the 80nm core.
It just kinda dumb that people keep repeating things are impossible that they misread from the TS.
ps.
all these new features are likely just to make a bigger gap perf. between the high-end and mid-end.
Till now the high-end always costed alot more then the mid-end but in some games the peformence gap wasnt always big enough to justify the price diffrence.
Wich makes me wonder what performence the mid-end will have asin pipes and bit bus.
1.5 seems hopefull as with the mention of a water Unit with the Card at the rumoured speeds but the reporting of some sites needs to whip their Audience up for their "certain" Reporting Skills to have their niche
whens this card around December or January?
Looking back, 50% increase in transistors number and 50% increase in clocks is plausible from one gen to the next, but there' more than 100% increase in both in this case. My take is unless those specs are way off, we're dealing with a two-chip GPU here.
so this is just more rumors? great more :spam: to sift through.
If the two chips are a reality it is nice to see nVidia pushing the bleeding edge forward with GPUs. I wonder if the midrange cards 8600GTX/GS will have a two GPU set up like thier higher end brethren.
An problem with one chip with two dies idea. The amount of memory chips needed for the design doesn't leave to much room for the rest of the circuitry, unless the card would be very long. Something like the 7950GX2 with 2 PCBs makes sense. The coGPU would be on the inside PCB with aircooling. The main GPU with more memory would be on the outside with more space for water cooling.
Fun Stuff indeed.
Well I had my own theory on it.
2 memory elements, 512mb 256bit, 256mb 128bit.
Considering they've basically said (from nearly every source I've read anyway) that it will be basically a DX9 card with DX10 extensions for DX10 compatibility. Ponder this.
Card is split for DX9 & DX10 (sort of)
The 512mb element of ram is used for the main section, not too dissimilar from current architecture with Pixel & Vertex shader processors, and a second unit for geometry shader processing, with its own vram module.
Second section only gets utilised in DX10 apps.
Seems plausable, plus considering they've said its a 384bit interface (even though it would actually be 256bit + 128bit seperately), so maybe the whole 700m transistors could hold some truth if the second DX10 core element took up maybe 1 third of those(as it would need a geometry unit, registers, memory controller etc)?
So in TOTAL, may be somewhere close to 700m, but its 2 seperate chips (similar to the dual core argument, though it would be 2 completely different cores for different tasks)
I agree with a lot of that as plausible...Most of us seem to agree on the two-chip solution as the most-likely possibility, as nn_step pointed out 700M on 80nm is crazy...It'd prolly be over 4cm2 if one die equalling a ton of heat. Your theory seems about as possible as any other if not more-so.Quote:
Originally Posted by Dublin_Gunner
The GT could then use a 64-bit (128mb) second interface, while still using a 256-bit original interface with 512mb of ram and still equal 640mb of total ram on a total 320-bit interface.
Wouldn't be surprised to see a 8/4 split memory config...Or both 10/12 configs on the flipside of the coin. If those rumors did answer some questions they surely opened up a ton of new ones.
if the 2nd core is for Physics and its performence is actualy good then Ageia is dead.Quote:
Originally Posted by LOE
And yeah 512Mb and 256Mb and 2 dies would take up alot of space but that doesnt mean they would need a really large card or 2 cards like the 7950GX2.
Just check out the new ATi cards with digital PWMs they need alot less space on the board.
It would be possible. 1 8800GTX contains 2xG80 cores. Thats what the boys at Beyond3D are predicting.Quote:
Originally Posted by zakelwe
So yeah the general consensus seems to be that we are missing a big peice of the puzzle, and thats only if these specs are real.
Someone start buying flowers for Ageia's grave
Dude, AGEIA was stillborn.
You make a good point LOE, the sucess of PPU would mean that tons of other comapnies would try to come up with similar innovative ideas. Forceing more and more cards on us.
It's logical. 48 pipes with unified pixel/vertex shaders. 2 complete ALUs per pipe. 96 total ALUs, or 12x the vertex shader ALU capability (8) and 2x the pixel shader ALU capability (48) of G71. Essentially 2xG71 with partially unified shaders instead of separate pixel shaders/vertex shaders and separate geometry shaders (16?) rather than vertex shaders. This would go along with the geometry shader being on the separate die using it's own memory bus and essentially not used in dx9. It'd be kinda like Xenos With a DX10 bolt-on.Quote:
Originally Posted by LOE
If that theory is correct, the 2x/16x number is deceiving market-speak because obviously you couldn't have either at maximum load at the same time gaming (you could use all ALUs while folding/doing other gpgpu things though which is sweet) but regardless makes sense. That 12x/2x part is the thing that makes me believe the specs are real as only marketing would use that obscure reference rather than the actual number. :)