And you need to stop getting your panties in a bunch.
Because SeekandDestroy suggested that it should be possible to design multi-GPU cards in such a way that they use the same memory. And I Was talking about that. Redoing the chip is one way. ROPs are decoupled in GPU pipelines these days and so, yes it is possible. But that's not what I was suggesting. If you thought it was then I'll apologise for my poor way of communicating it to you. Rather, I was suggesting going with the 768MB cards because if they're doing a shared memory version of it that is all that's really necessary for a twin 460 card to be great.
Remaking the chips is usually a bad idea because of the high initial cost of ASICs. Since this is entirely a digital circuit it is doable on FPGAs instead for small volumes, but that would still make it pricier than existing chips that are ASIC. I'm speaking from experience. Some years ago I designed an experimental processor, for real-time image processing and for nonlinear robust control, that had to go on ASIC instead of FPGA because even though volume was small it was a mixed signal circuit. that's was a different issue entirely because I work with analogue computation which is far far more powerful than digital. and this was a mixed-signal circuit done in VHDL with CMOS floor plans for the digital aspect, so only ASIC would do.
What I was talking about instead is using two 460s with 768MB 24 ROPs each for a 48 ROP card with shared 1.5GB memory, which would really be ideal for virtually anyone who wants a dual GPU card in that price range. It is highly likely you can do that with external control circuitry without modifying the chips themselves. Of cousre it is impossible to tell precisely without looking at the card design itself whether it would be feasible, but to me it looks like it should be possible. Nothing ground breaking needs to be done here either. I can see several ways they may set about doing soemthing like that.
If the whole point of your rant was to say that such a card is not economically feasible, then noted. I cannot agree or disagree unless I'd actually studied the designs and looked at how costly it would be, but I have a notion there is a fairly simple solution to doing his in practice. If it can't be done fairly easily then it makes sense to just stick to hte 1GB cards. or go with 2GB cards and have 4GB. But I often in engineering for something like this where you're dealing with essentially the same data on two buffers there are simpler ways of achieving this. For example, connect circuitry that fools one GPU into thinking that its fetch requests are being satisfied, and then feed it data from the same memory buffer when it makes a request. Again, impossible to say if that will work without actually looking at the designs.
But unlike you I am not going to knock an idea based on some superficial information I read off random review sites who are written mostly by people who aren't even microelectronics engineers themselves.
Someone bumped this and I only just saw this, literally everything you are saying is wrong to such a high degree its laughable.
sli doesn't work with two cores on the same frame, each core works on a different frame and will need DIFFERENT data, most of it will be the same, but any differences and "copying" memory from one to the other would fail, completely. You can not share memory, this is the way Sli/xfire work, they way they've worked for what, 7-8 years now and the way it will continue to work for some time.
remaking the chips using FPGA would be lower than a new ASIC but higher than an old one, suggesting it is truly ridiculous.
You somehow think that designing an entirely new chip, and sourcing production from somewhere completely different outside of their normal manufacturing chain, and validating, testing, warrantying a completely different type of chip would be viable, at all? All because you think saving a few bucks on a few extra memory chips is a good idea, hundreds, infact, easily thousands of engineer/design work hours, to make something more expensive, that wouldn't work nearly aswell, that almost certainly would have massively stunted clock speeds, to not spend $3-4 extra per chip on 6 chips.
As pointed out there are 2gb cards, how do they do this, double density chips, so 6x 256mb chips instead of 6x128mb chips, they cost next to nothing, you want them to design a memory sharing device(which would not work with their current drivers/sli method) just to avoid about $15 in extra memory? You're talking about anywhere from 100k's of dollars, to millions, to avoid a $15 cost and provide a lower performance solution.
24 rops would be more than acceptable on a card with that level of performance, its sli, you've moved up from 460gtx performance to well behind 480gtx level performance, you'll be buying a card like that for 1920x1080 AND UPWARDS, not midrange resolutions, you want to use 768mb per chip, thats 6 chips, and 6x32bit memory bus vs using 1gb per chip, which would be 8 chips or 8x32bit bus, you think 24/48 rops would serve better at a higher resolution and higher graphic settings that are intended for sli/xfire setups, despite 24 to 32rops having a noticeable performance on SINGLE 460gtx's.
EVery one of your idea's would for massive, massive expense, and a massive change in sli working, and a massive effort for driver support(because it would be the single card since the 8800gts to HAVE to work completely differently to the rest of the Nvidia range), for years because it would need support at the very least till the warranty was done, all to save a few bucks on memory chips.
LIke I said before, everything you suggested was insanely daft, for many many many different reasons.
Why it would cost 100k's of dollars or more is, you're talking about the entire driver team devoting time to a new driver, for YEARS, you're talking about designing a new external memory sharing device, and pcb's, and validating, and production. You're talking about dozens of people working for months on end, when you have the chips sitting right there, ready and waiting. While your idea will drastically reduce performance, for no earthly reason, all because you think the more gpu power you have, going for 2x 460gtx cores, you really don't need "that" many rops, or that much bandwidth for that kind of resolution.
1gb cards have higher bandwidth than 768mb cards, 1gb cards have more memory, every 768mb card has a core that physically HAS 32 rops on it, and yields at this point that are ample, so you'd disable 25% of them, just for the sake of it, then you'd disable 25% of its bandwidth for no reason, and then you'd expect people to buy a card for high resolution thats faster than a 480gtx, but hasn't got the raw horsepower anymore.
DO you know why Nvidia put two FULL 580gtx's on the 590gtx, and they didn't randomly disable 25% of its rops, and 25% of its bandwidth, and spend millions on a "new sli" for a product that will sell less than 10k units, because its daft. Do you know why AMD didn't do that either, because its just very very daft.
You have a core, working and ready to go, why would anyone purposefully stunt it?
