Nvidia next generation chip will have dual core

[AthlonXP1800]

I think G70 and G80 are very mystery chipsets either former NV50 or top secret name for NV47 or whatever, I guess here not much new stuff in it but could be like what it did from R380 to R420 with not much new stuff in it. G70 could be bug fixed dual core NV40 chipset manufactured by IBM 0.09 micron technology. It may seem ATI had missed dual core boat this time. :)

http://www.theinquirer.net/?article=21822
http://www.theinquirer.net/?article=21827

[Remi]

Huh?? Multicore GPUs?

The concept is rather strange to me, because when CPUs are taking the multi-core road they're kind of imitating GPUs. GPUs are already inherently parallel, so multi-coring them to parallelize them doesn't make much sense to me - GPUs are already parallel!

I can see a few problems that multicoring them would cause in the design, but the only advantage I can see is that it could improve the yields, if each core can be build separately and then assembled in the chips case. Which seems to me like a rather costly operation, making the whole idea not very cost-effective. I dunno...

Maybe assembling cores is in fact a cheap operation?

[Smokey]

Havent you guys seen this?

http://pcmoddingmy.com/comment.php?comment.news.777


http://lp.pcmoddingmy.com/albums/userpics/Misc/6800udual.jpg

More pictures HERE (http://www.hardspell.com/news/showcont.asp?news_id=11588)

[jolle]

yeah that sounds wierd..
Why not make a ex. 32 pipeline GPU instead of cramming 2x 16 pipe cores on 1 chip? it would be pretty much the same thing one would think..
I guess it might be easier to do f.ex a 16 pipe config and put 2 cores together, then it is to produce 1 core with 32 pipes, since there are 16 pipe GPUs but no 32 pipe ones hehe, in which case it might make sense..
If they could be "fully" linked up so they would act and perform like a 1x 32 GPU, it might be neat, would they also have their own dedicated memorybusses it would do wonders for bandwidth, 2x 256 or perhaps even 512bit busses..
the later with LMA that would be.. 8x64bit capable to up to 2x512bit right?
compared to NV40 4x64bit

Ofcource, coming from the Inq it involves a heap o salt to be taken during the read..
Wonder if this fuad guy publishes things based on "what some guy just said in a anonumous email", plain guessing or if there is any substance to it..
Who knows, maybe it is true.. time will, as always, tell..

Havent you guys seen this?

http://pcmoddingmy.com/comment.php?comment.news.777

That would be DUAL GPU, not dual core..
Dual core would be 1 GPU containing 2 GPU Cores..

[AthlonXP1800]

Maybe assembling cores is in fact a cheap operation?

Yes :)

I think here are something very big than these G70 and G80, you see Nvidia joined Sony 3 years ago to developed custom NV5A chipset and about the same time Nvidia joined IBM to make chips for them. IBM and Sony are the partners to develop Cell CPU, it seem Nvidia had the deal as part of the whole package to grant full access to study Cell CPU technology. I think that what Nvidia got borrow ideas or part of Cell CPU technology integrated into G70 and G80 chips make it much cheaper to manufacturing, much easy to develop and design multi core GPU. NV4x chips are full programmable GPUs, I think it behave similar as the CPU does.

[Smokey]

yeah that sounds wierd..

That would be DUAL GPU, not dual core..
Dual core would be 1 GPU containing 2 GPU Cores..

My bad, acctually read the post above mine incorrectly :o

[MUYA]

that would be 2 GPU cores on one die :)

[jolle]

My bad, acctually read the post above mine incorrectly :o
I do that all the time hehe,Im the worst skimmer ever..

[GigaDrive]

wild guess:

Nvidia G70 = NV4x with 20-24 pipes. ( NV47 ? )

Nvidia G80 = NV4x with 24-32 pipes. ( NV48 ? )



:D


I do not believe that G70 and G80 are NV50 or NV5x. I believe G70 and G80 are both refreshes of NV40, thus NV4x GPUs.

....ala NV25 and NV28, both GeForce4s, refreshes over the GeForce3|NV20.

maybe at least one of them (G70, G80) is dual core if not both.


*imagines dual-core GPUs used in 2 cards in SLI*

:eek: :D

[zoomy942]

Yes :)

I think here are something very big than these G70 and G80, you see Nvidia joined Sony 3 years ago to developed custom NV5A chipset and about the same time Nvidia joined IBM to make chips for them. IBM and Sony are the partners to develop Cell CPU, it seem Nvidia had the deal as part of the whole package to grant full access to study Cell CPU technology. I think that what Nvidia got borrow ideas or part of Cell CPU technology integrated into G70 and G80 chips make it much cheaper to manufacturing, much easy to develop and design multi core GPU. NV4x chips are full programmable GPUs, I think it behave similar as the CPU does.


the most exciting thing about nvidia's next step... is that we know nothing. i havent seen a company this closed up about their next product. the usually leak a little something. i have a feeling they are going to have something that will be very diferent from what we normally think of as a GPU. and they are waiting for ATI's new thing.. a nd then they are going to unveil it. with a fully programmable GPU, imagine SLI with GPU's that are created to interact with one another from the start. i imagine it kind of like 2 NIC's. they are made from the beginning to expect each other and interact in the most efficient way, but if there is only one, thats okay too.

[Remi]

I think that what Nvidia got borrow ideas or part of Cell CPU technology integrated into G70 and G80 chips make it much cheaper to manufacturing, much easy to develop and design multi core GPU.
Well, while that may seems a good idea, one have to remember that Cell processors are streaming processors, they're somewhere between CPUs and GPUs, which makes them... ...less effective than a GPU at doing GPU work. (Deano Calver already explained this a bit here (http://www.beyond3d.com/forum/viewtopic.php?t=20447)).

NV4x chips are full programmable GPUs, I think it behave similar as the CPU does.They remain fundamentaly different, because of the way they handle RAM. To a processor, normal RAM (DDR) is very slow, and without mechanisms to hide that, our processors would be iddle more than 90% of the time.

CPUs don't have much choice, they must use 2 levels of cache, because they can't plan in advance when and where in memory they'll need to access. Only the programmer who wrote the program running could know that, and he would require to know the detailled context in which his program is running in order to say something.

GPUs on the other hand can forecast much better their needs to access the memory because their work is much more regular and predictible than the execution of a CPU program. That makes it possible to design the chip with very small caches. For instance the vertice cache's capacity count is not in mega-vertices nor in kilo-vertices but in vertices. Waaaaay less than the caches of a CPU.

As new capabilities are added to the GPUs, they do look a bit more like CPUs, but only a little bit. They still lack very important CPUs capabilites such as writing anywhere in memory - and that's precisely because they don't do it, unlike CPUs, that they can be that fast at shading...

There is no doubt that nVidia knows a lot already about the Cell, but what lessons they have learned from it that is applicable to the GPUs remains very undisclosed...


Edit: I just read again Deano Calver's post, and I noticed again something which left me perplexed when I read it first (and I kind of forgot it since then :D ) : "GPUs will ship with multiple 'units' as illustrated here just as Cell ships with multiple SPUs".

Interesting, to say the least...

I think that what Nvidia got borrow ideas or part of Cell CPU technology integrated into G70 and G80 chips make it much cheaper to manufacturing, much easy to develop and design multi core GPU.Looks like it might be worth to think about it twice!

But the concept still looks weird to me...


Edit 03-17: I'm in fact begining to wonder... with multicore CPUs, parallelism is a concept that is getting some visibility through the "multicore" term. What if quads were suddenly relabeled "core"? Well, then one could say a (high-end) graphic cards are having 4 cores (and probably 6 soon), twice / three time the number cores in a CPU! That would be usefull to let the general public know that GPUs are "more powerful" (more gflops) than cpus... Technically debatable, but... That's one possibility.

[technews]

multicore is the way to go...

so think about cards with two dual-core cpus each ... like the Gigabyte...
and then add another card on SLI

that will be 8 cores working together! heheh

I wonder :)

[Remi]

Just a little post in order to keep this thread updated. Please remember this is still only speculation.

The very first idea is often the good one, and I tend now to believe that could applies here.

I was assuming that in order to connect two dies in the same chip case, one would proceed as it was done to connect the chips to the pins (and maybe still is, my little knowledge in this field isn't up-to-date) - with a machine that positions and welds thin wires between them.

It looks likes that will soon be passé however, as small circuits pieces can now be used instead to make these contacts from one die to another. That's clearly not used today to produce the Cell processor (Cell is one die), nor the actual dual-core Pentium (the Smithfield core is one die too); but it has been researched at least by Sun and it's rumored that Intels plans to produce multi-die multicore pentiums. Which is why I assume that the multi-die connection technology is mature enough (or close to) for production.

That would have a few advantages. If you split your monolithic circuit into several dies, you also split the risk associated to the yields (one failure means only a die fails, rather than a whole circuit).
- that increase the return on investment (less circuit loose).
- that increase the volume of production of your high-end products (it's easier to assemble small good dies to make a high-end circuit).

I would not really bet that it will lower much the prices of high-end cards, as it wouldn't reduce the die area (on the contrary, it would increase a bit due to the additional communication circuits).

Although using the inherent flexibility seems good (putting a different number of dies inside various chips to achieve various price and performance points in the product range), that looks rather difficult to achieve without significant costs in die area (too many communication units required). Unification of shader hardware could help to enable a multi-core vision (a core could be four shader units) with only a few dies (hence less additional costs due to communcation units). Other possibilities might be possible, but load balancing accross the dies would probably be more difficult.

The idea doesn't really look stupid, in fact it even looks worth of investigation...

But all that is just speculation, and even worst, it's speculation on hardware coming from a software guy... :D

[MustangSVT]

so think about cards with two dual-core cpus each ... like the Gigabyte...
and then add another card on SLI
What are you talkinga bout? What gigabyte?

Dual core is two GPU cores on ONE chip. This has never been done before. The Gigabyte 3D1 card or that ASUS card simply have 2 separate CHIPS/DIES, they do not have dual core. All these cards do is put 2 chips on one card and work them together through nVIDIA's SLI which is not 100% efficient. Dual core would mean two GPU's on die and them working much more closely together and in a different way than just simply like SLI.

Anyone can put on 2 chips on one card if they really want to. 3dfx did it with the V5 5500 5 years ago, and with the 6000, they put 4 chips on it, all through their SLI. There's been professional ATI cards with like 2 or multi 9700/9800 chips on them for like two years now. It's not anything new.

Now putting two GPU cores on one die to make them work together in a much much better way than SLI hasn't been attempted by nVIDIA or ATI so far, and it probably won't for a long time from now. The next G70 or G80 or whatever will not have a dual core GPU on it.

[latinmike]

Havent you guys seen this?

http://pcmoddingmy.com/comment.php?comment.news.777


http://lp.pcmoddingmy.com/albums/userpics/Misc/6800udual.jpg

More pictures HERE (http://www.hardspell.com/news/showcont.asp?news_id=11588)

bad link try again :cool: