News First AMD Torrenza partners come out of the woodwork

[WilHarris]

http://www.bit-tech.net/news/2006/09...of_the_woodwo/

[DougEdey]

Soooooo want this to be the physx successor. I thinK ATi could make an addin card that'll fit and spread the load!

[EQC]

I'd like to see Intel license hypertransport...then I could build myself a computer with one AMD processor and one Intel processor. End all that "____ is better" nonsense for good.


Also: anybody know how RAM resources will work with this sort of platform? Will each socket have it's own associated RAM slots...or will both processors share system RAM? While there'd be some advantage to sharing system RAM, it is much slower than high end graphics RAM and other specifications...

[Lazlow]

Ooooh, this sounds very interesting. It opens up a huge amount of possibilities.

[Da Dego]

Originally Posted by EQC

Also: anybody know how RAM resources will work with this sort of platform? Will each socket have it's own associated RAM slots...or will both processors share system RAM? While there'd be some advantage to sharing system RAM, it is much slower than high end graphics RAM and other specifications...

As far as I've come to understand the Torrenza platform, it will share system ram - hence a big part of why it needs to be on the HTT. Whether or not chips could come with an on-board cache to help buffer that would probably depend on the specifics of each co-processor that would be implemented. For right now, think along the lines of back when we needed math co's...they didn't even require seperate memory, just took portions of instruction blocks that were already being allocated to the CPU.

It should be interesting to see where this shakes out - for as much as we keep trying to pull stuff off the chip, there's a whole second group trying to pull stuff back on it.

[Tyinsar]

Originally Posted by Article

Eventually, AMD believes that co-processors could land on the desktop to handle physics and graphics calculations - part of the reason for its purchase of ATI.

This would mean faster communication with the CPU & GPU but, as others have pointed out, some GPUs use faster ram. I'm waiting to see how this would balance out. It will also be interesting if they can get Nvidia on board (pun intended?) - If not then I think this will die quickly (on the mass consumer end anyway) unless ATI's chips come in with decent cost savings (tricky with extra R&D costs) and equal or better performance (using only DDR2 shared over HTT).

Originally Posted by Da Dego

...It should be interesting to see where this shakes out - for as much as we keep trying to pull stuff off the chip, there's a whole second group trying to pull stuff back on it.

QFT, this was also my first thought when I read this.

[Tulatin]

the way this is penned, Torrenza seems to be built to utilize the current crop of 940 mobos - and considering that these are boards intended for independant dual computing, it means that each chip will have it's own bank of ram, juts like the current batch of opterons.

[Neo.Luxembourg]

This is realy intresting: http://www.reghardware.co.uk/2006/09/19/ati_gpgpu/

If you think about it, an what can be done for computers that do like Encoding or Calculating the same Algo all the time around ....

[Tyinsar]

Neo.Luxembourg: Awesome link , thanks

[Lowsidex2]

What's with the cycle of life? Wasn't this common back in the 486 days?

First it's divide and specialize. Then it's simplify and consolidate. Now we're going back. CPU, Co processor, GPU, AI processor, Physics Processor. In a couple years we'll be back to using raw horsepower of new chip to handle it all again.

[bluebox]

"...Called 'RoadRunner', it will see 16,000 Opterons paired up with 16,000 Cell processors in the 940 form factor, communicating on dual-processor motherboards. The Cell processors will handle computation, whilst the Opertons will handle I/O..."

Just wanted to point out that this isn't right. The Roadrunner cluster is not using Cell and Opteron in the same servers. The Opteron processors are in x3755 rack servers and the Cell processors are in QS20 blades.

http://www-03.ibm.com/industries/gov...?g_type=hpfeat

[Sparrowhawk]

Personally, what I would like to see is co-processors on a home level. Imagine this: boot your OS of choice, using all processors. While doing mundane tasks you use the smaller co-processors and the big main only as needed, and for gaming/rendering/etc the main processor does all the work, while the co-processors do system tasks and/or keep the OS stable.

Granted, on the server side, this should be nice this co-processor scheme.

[r0z|3o0n]

This reminds me disappointingly of the math co-processor days of when I had my 386/486. Either it will turn out to be a really cool initiative that opens new frontiers in processing, or it could turn out to be the only way AMD can keep up with the power of the new intels - by adding, a-la-386, a co-processor.

If it's the latter it will be sad, because the market will swing back to Intel domination. Apart from bleeding edge people don't want to pay for 2 processors when one can do the same job only SLIGHTLY slower. AMD's competition has been a fantastic thing for the consumer - I can buy a 2.66ghz dual core for less than $200NZ but if Intel gets a jump on AMD I can see processor prices climbing again. Anyway, i'm rambling, and i'm at work, so I should probably get back to doing something constructive

[automagsrock]

Originally Posted by the article

"...Called 'RoadRunner', it will see 16,000 Opterons paired up with 16,000 Cell processors in the 940 form factor, communicating on dual-processor motherboards. The Cell processors will handle computation, whilst the Opertons will handle I/O..."

Boy I would really like to have that in my basemesnt just to show off.


I think this new co-processor stuff is really cool. If they can get it to the personal computer level for cheap and prove that it is powerful, I think AMD will have a winner.

[Flibblebot]

Can't really see it becoming common in mainstream PCs though - if you're a non-AMD company, you'd be missing out on a huge potential market because those people use Intel instead of AMD.

Even if Intel did introduce something similar, they wouldn't use AMD's socket format, they'd use Intel's socket format. This means that, as a third party, you'd end up having to design each of your products in two separate packages (with all the associated differences in bus architecture etc) to get access to all markets.

The reason PCI, PCIe & AGP worked is because they were standard connectors. Having a proprietary connector is a backward step, regardless of the possibilities.

[Marquee]

This is a step in the right direction for AMD. I love the idea of having CELL CPU paired up with a Opty to work together. That will be a crazy computer. Can you imagin me getting my hands on a cell CPU and paring it with a Opty for home use.

[Tyinsar]

Bluebox: thanks for the clarification & welcome to the forums

Originally Posted by Flibblebot

...if you're a non-AMD company, you'd be missing out on a huge potential market because those people use Intel instead of AMD...

The Intel platform does not have HTT - they've done great stuff with the cache on the C2D but their FSB still has way less bandwidth - Their performance will struggle with more chips unless they fix this. Also - if AMD/ATI can produce a socketed general purpose GPU (which could also do physics and be used for other functions) that is as good or better than Nvidia's at a cost savings then Nvidia would almost have to get on board or lose any chance of sales on AMD systems.