Multi core GPU, the future?

[alg7_munif]

Multi core CPU is not something new today and we already have the SLi and Crossfire technology for quite some time. There are also some single slot dual cards like the 7950GX2 and Sapphire X1950Pro Dual already out there, GeCUBE Dual HD2600XT will join the club later.

After some reading, I must say that I'm quite impress with the low power consumption of the new 65nm HD 2000 series from AMD and using 65nm process also means that the GPU die will be smaller. I'm sure that nVidia will also usa a smaller process in the near future. With a lower power consumption and a smaller die, I think that it is possible to have more than one GPU on a single card already. My question is, will the GPU technology follow the foot step of the CPU technology by using multi core processor in the future? The only drawback that I see is you need a Crossfire/SLi mobo to such a card from either company. Hopefully they will stop limiting multi GPU technology on certain chipset in the future.

 

[meatfestival]

Multi-core GPUs don't make a whole lot of sense. GPUs are already massively parallel, therefore they just need to add more pipelines or stream processors or whatever.

Multi-GPU cards are a slightly different kettle of fish.

 

[Karma]

Multi-GPU cards are a slightly different kettle of fish.

Yes, but ATi's R700 is supposed to be several smaller GPU's interlinked together to perform larger operations (on the same package), somewhere near the teraflop range. nVidia's G90 is also supposed to be in that processing range, but no one knows the general layout of the chip.

I suppose you could see the dual chip cards as the future, but only on a philosophical level.

 

[DrFreeze]

Multi-core GPUs don't make a whole lot of sense. GPUs are already massively parallel, therefore they just need to add more pipelines or stream processors or whatever.

Multi-GPU cards are a slightly different kettle of fish.

exactly, especially an architecture like R600 or G80, which is essence is just a whole bunch of floating point specific cpu cores bunched up together (scalar in the g80 case, superscalar in r600)

the only reason to use multi GPU cards is to circumvent die size/yield limitations

 

[alg7_munif]

I wonder, which is more cost effective, designing a new chip that have double the pipeline or using two existing chips to gain about 75% of the performance. Manufacturer can design a new chip with more pipeline for every new generation cycle but I think that it would also be more profitable if they also make a double core card from the same chip. Let say that you want to have the best graphic solution of current generation card, would you buy two 8800 Ultras in two cards or a single card with two 8800 Ultra core? I know that using two Ultra core on a single card would push the limit of the PSU but what if the power consumption and die size of the Ultra chip is like the HD 2600XT chip?

 

[legcramp]

Yes, but ATi's R700 is supposed to be several smaller GPU's interlinked together to perform larger operations (on the same package), somewhere near the teraflop range. nVidia's G90 is also supposed to be in that processing range, but no one knows the general layout of the chip.

I suppose you could see the dual chip cards as the future, but only on a philosophical level.

Ati needs to learn how to correctly make a single gpu core work properly first

 

[Karma]

Ati needs to learn how to correctly make a single gpu core work properly first

Heh. Yeah, but that's not the topic.

I wonder, which is more cost effective, designing a new chip that have double the pipeline or using two existing chips to gain about 75% of the performance. Manufacturer can design a new chip with more pipeline for every new generation cycle but I think that it would also be more profitable if they also make a double core card from the same chip. Let say that you want to have the best graphic solution of current generation card, would you buy two 8800 Ultras in two cards or a single card with two 8800 Ultra core? I know that using two Ultra core on a single card would push the limit of the PSU but what if the power consumption and die size of the Ultra chip is like the HD 2600XT chip?

It is definitely more cost effective to have smaller, inter-scalable, chips linked on the same physical package. This allows the engineers to lower complexity, lower defect rates, increased speeds, increased usable chips per wafer, etc. So far everyone has been designing ever increasing dies with increasing performance potential (and power draw), but also increasing engineering difficulty, defect, and failure rates (think R600). You can only have so many working chips per wafer.

Of course, GPU's are already massively parallel. But if you were able to link some existing chips together for increased performance without having to design a while new architecture, life would be so much easier. SLI/Crossfire is an example. In your example you give 75% improvement. But in Multi GPU systems you can already get up to a 100% improvement by adding a second linked card. In theory you could get the same improvement by adding a second GPU to the same card/PCB (although in practice so far they haven't been much more than something to buy for bragging rights).

Where the real future lies is not GPU's linked across different cards (SLI/Crossfire), or even GPU's on the same card (X1950Pro Duo). Instead expect smaller, more independent, but massively scalable GPU cores, all put together on the same die. This will make it much, much easier for the designing company to stick 1,2,3,4,5,6,7,8,9,10......32......64.....128.... as many cores as they want onto one PCB. It will also allow much more economic grading of graphics card. Instead of low, mid, high, & enthusiast class of cards, imagine Intel (and I say Intel because they are already doing this exact thing with their "Larrabee" project in 2009-2010) releasing one card with two cores for the ultra low segment, then another card with 3 cores, then another card with 4 cores, all the way up to cards with 32 cores on the same card!

Larrabee: 16 Cores, 2GHz, 150W, and more...

Pretty cool, huh?