Next Gen cards: Multi Chip?

[Evildead666]

There is lots of speculation concerning the next Nvidia chips (NV30) saying that this should be the version which includes 3dFX hardware enhancements.
While it is true that the GeForce lineup are basically reworks of the first GeForce chip, on smaller processes and with the a few enhancements along the line to follow DirectX.

This next one HAS been touted as being totally different.

A multi chip solution has been earmarked before (by 3dFx) and they were on to something they believed to be powerful (Rampage and Sage).

Are mulit chip solutions the only way forward?

Just remember that the fastest and easiest way to solve a problem is to Divide and Conquer.
Split the problem up into smaller, easier to handle problems.
Hence the multichip: One controller, multiple gpu's.

The memory interface is hectic, and the texture memory is bad at the best. Do you have large memory with a copy of texture for each chip (in their own memory) or do you use a UMA style architechture, where the memory on the board is shared by all the chips (i suspect a V large bandwidth needed for say, 4 chips sharing the same memory...4x what one chip would need?)

Multi chip is the way to go.
what do others think?

[piaxVirus]

I don't think multiple chips are necessary.

Matrox and ATI will not be releasing any multi chip cards. Nvidia, most likely will release some mutli chip cards as high-end consumer cards (quatro series). The only reason they are making the NV30 multi chip capable is because they need to replace the aging VSA-100 chip that Quantum3D uses in their commercial cards.

[Drizzt]

I think that more than multichip, an interesting future would be multi-die.

[Wombat]

Same thing. Multi-core?

[rugger]

Nv ins't going multichip

I hate RUMOURS

Hate Hate Hate Hate Hate

If I had a dollar for every false rumour in the graphics industry, I could go buy a Parhellia or two

Nvidia isn't going to multichip, simply because it is much more expensive to produce.

They may produce one or two very high end cards with multichip, but with their current stance on multichip solutions, I don't expect it.

One more thing, everybody says that 3dfx were working on a "fantasic new chipset" when they died and were brought out by Nvidia. I am just questioning:

1) Just HOW fantastic was this new chipset? Was it as *fantastic* as the Parhellia was, or the Radeon 8500 was. (very heavy sarcasm here) It could have easily been a very mediocore chipset, like the VSA100, or simply just an adequate chipset, like the Radeon and the Parhellia.

2) How complete was this design?

Nvidia probably only got some engineers and patents for their trouble buying 3dfx. I suspect quite strongly that the NV30 will not be the next 3dfx chip, as much as 3dfx fans would like it. I strongly believe that the new NV30 will instead be a classic Nvidia chipset, with a few optimizations / features brought in by 3dfx. Multichip may be part of it, but most of the 3dfx stuff will be where we don't see it

[Buback]

I can understand how a multichip solution would cost more to research. But what about extending the life of a chip design by adding another chip. obviously it would cost something to redesign, but it might cost less than a whole chip redesign.

[Evildead666]

From what i haev read Moores law is slowing down.

This means that the speed and size of chips isn't going as quickly as before (past 27 years).

This could possible mean that the chip speeds are gonna increase slowly, as the complexity increases.

Both ATi and 3dFX have used multi chip technology, to rightly increase the bandwidth and rendering power of the cards.

Look @ PC's and Macs, you get a dual processor solution when u have a lot of very complex number cruching to do, so logically, when you have a lot of complex graphics to do, then u use a multi chip graphics card.

Ok, the cpu doesn't treat the same data, and not even in the same way, but its still useful for comparison.

U could have one chip solution, which can do everything in OK quality,
or you could have a multi chip solution that divides the problem, and sends the data to the chip that has to process it. Each chip takes care of its job and sends the data back for display.

It may be more complex, but its more powerful.

I also beleive that nvidia will only introduce the next gen chip in multi chip for CAD and rendering purposes etc...
But this also means that they have the possibility of doing the same with gamers cards, when the chips get into their 2nd wind (change burn from 0.13 to 0.009 for example) increased yield, and lower cost cores could spell a lot of good multi chip cards..

[Pace]

Re: Nv ins't going multichip

Originally posted by rugger
I hate RUMOURS....Nvidia isn't going to multichip, simply because it is much more expensive to produce...They may produce one or two very high end cards with multichip, but with their current stance on multichip solutions, I don't expect it...says that 3dfx were working on a "fantasic new chipset" when they died and were brought out by Nvidia...HOW fantastic was this new chipset? ...could have easily been a very mediocore chipset, like the VSA100, or simply just an adequate chipset, like the Radeon and the Parhellia.....Nvidia probably only got some engineers and patents for their trouble buying 3dfx. I suspect quite strongly that the NV30 will not be the next 3dfx chip, as much as 3dfx fans would like it. I strongly believe that the new NV30 will instead be a classic Nvidia chipset, with a few optimizations / features brought in by 3dfx. Multichip may be part of it, but most of the 3dfx stuff will be where we don't see it

Glad you dispelled all those rumours with your facts

[Drizzt]

It's obvious that a graphic card has the best use for a multichip solution, 'cause it's SIMD work that the card does.

However, I don't think that a multichip solution will be the must in the next month. As I wrote a few above, I think that an interesting and immediate solution for improving performance is a multi die: imagine to subdivide the chip into functional parts, and build each of them on a different die; this will help reducing heat produced by having an huge number of transistors in a minimum space, spreading the heat over a larger surface.
Thus, you can higher the chip clock...

[Ssin]

I don't think that producing a multi chip solution would be more expensive for a company like nVidia. Of course they would have to design something to accept 2/4/6 chips on the sames cards but.... think about it, nVidia only produces chips, they do NOT manufacture cards. I think it's really a smart thing to do. If I were a card manufacturer and I wanted to have alot of different cards models I would just have to add a chip to a card that is designed to accept up to say 6 chips. That for the same price tag as for a single chip card... I guess that's what I would do if it was possible and if nVidia was doing that....and of course I were a card manufacturer.

BTW what does it take to be murcer of supermurcer?

[MetalCartman]

>=100 posts: MURCer
>=1000 posts: Super MURCer

[Wombat]

Moore's Law isn't going anywhere any time soon.

[Cobos]

Well there is a pretty definitive limit to how small you can make transistors... I dare say when you reach around 0,5 nm you are at the limit as that is about 2x the atomic distances in most metals... And that is pure metals... The closest you can pack metal atoms is about 0,1-0,15 nm apart... and with 130nm the upcoming distance it isn't that much more to go further...
Of course if quantum computers or similar "wierd" designs actually get producable then all bets are off...

Cobos

[Wombat]

No, thank you, there is no definitive limit. We'll find our ways. But the real issue isn't gates anyway, it's wiring.

[William]

You sound like you have a specific solution in mind Wombat...