So whats NEXT for ATI?

[TheBluePill]

ATI has a real winner with the 3870 X2, but that crown (which is arguable at best) isnt going to last very long if nVidia's newer products perform to expectation..

So what is "next" for ATI? They cant afford to rest.

 

[Pilot53]

Idk about a real winner, turn on aa and see what happens. But I too am interested to see what ati has planned next. What they really need is a whole new architecture that doesnt need 4 gpus to beat an nvidia gpu. Any news on when they are going to have something totally new, or at least what their next planned release is?

 

[TheBluePill]

Idk about a real winner, turn on aa and see what happens. But I too am interested to see what ati has planned next. What they really need is a whole new architecture that doesnt need 4 gpus to beat an nvidia gpu. Any news on when they are going to have something totally new, or at least what their next planned release is?

I really like the Idea of Going "multi-gpu" in mainstream products. If they start relasing several "X2" type cards with each generation, developers will more widley support Crossfire and the whole ballgame will change, even for SLI, since nVidia will have to compete. (a 3850x2 for $279 would be great!)

its all VERY good for the consumer.

 

[Jakalwarrior]

Multi anything always has bad scaling but if they want to keep pushing... they can only go so far with each new process design. Im sure ATI should have some growing room though since they just had a die shrink but are still using pretty much the same core design. Course... those little dies make more money...

 

[Silus]

ATI has a real winner with the 3870 X2, but that crown (which is arguable at best) isnt going to last very long if nVidia's newer products perform to expectation..

So what is "next" for ATI? They cant afford to rest.

I think you already know what's next. There's a topic about it, right here in the first page of the ATI sub-forum. R700 is next and it should have a similar design as the X2, but with probably more GPUs.

Honestly, I would prefer that ATI actually developed a powerful single GPU, because, as was proven by NVIDIA with their first GX2 and now ATI with the X2, these multi GPU solutions rely too much on driver support, which is lacking for the most part. I don't exactly know what to expect from R700. ATI needs a winning product, but I'm not sure that going multi GPU, is a good way to reach the top. We'll see.

 

[Mr. Bluntman]

I think you already know what's next. There's a topic about it, right here in the first page of the ATI sub-forum. R700 is next and it should have a similar design as the X2, but with probably more GPUs.

Honestly, I would prefer that ATI actually developed a powerful single GPU, because, as was proven by NVIDIA with their first GX2 and now ATI with the X2, these multi GPU solutions rely too much on driver support, which is lacking for the most part. I don't exactly know what to expect from R700. ATI needs a winning product, but I'm not sure that going multi GPU, is a good way to reach the top. We'll see.

Yes, R700 is supposedly multi-core, and is speculated to be a single package, multi-die affair (multi-chip module). While I myself question the path they're taking , ATi seems very confident that this is the way to go forward.

 

[sam0t]

Single package with multiple cores glued together ala Intel Quad is the way to go imho. Intresting times.

 

[Mr. Bluntman]

Single package with multiple cores glued together ala Intel Quad is the way to go imho. Intresting times.

You can bin them for different range products, less transistors in a given area means less defects, and costs less to manufacture. A "win-win-win" situation, so to speak. Now if AMD would have taken that approach and just designed Kuma in the first place...

 

[TheBluePill]

AMD did the same thing with the intro of the X2 Processors. Changed the whole game and even intel had to play catch-up for a while.

Hmmmm.. Wonder if we will see a X4 GPU?...

Now.. if they just had a way to make them scale better...

 

[Kaldskryke]

Disclaimer: I'm not an expert, I have no sources. Believe me or don't.

Scaling isn't just a driver issue.

GPUs are already massively parallel. With CPUs, adding that second core is a massive change in how things are done (serial to parallel). But GPUs are already parallel where it counts, so doubling up on cores isn't going to change the game in the same way.

Notice I said, "parallel where it counts". There are parts of the "Rendering Pipeline" that don't benefit from parallelization. Adding more cores means all those parts become redundant. Then there are also communication issues between cores and all that so they can work together.

The fact is that multiple-core GPUs will never have 100% scaling, ever. You can stop dreaming of that right now. ATI isn't aiming for multi-core because it makes the best sense in raw performance. It's aiming for multi-core because of all the other benefits.

- smaller dies mean more dies per wafer and less cost.
- independent cores can be deactivated for power saving
- multiple SKUs can be created from the same core design = less cost.

Let's imagine a hypothetical single-core R700-A that's got a whopping 1.5 billion transistors and is rather large. It would probably perform quite well, there's no driver or scaling issues, etc, but it's big, low yields, expensive. That cost gets passed on to the consumer and you're shelling out $800 dollars.
Instead you could have a 4-core R700-B that's got about 2 billion transistors (500 million each). It would a bigger overall package and would consume more power at full load, but would perform about the same as R700-A. It's idle power draw would be lower (disable 3 cores) and it could quite likely cost ATI a lot less to produce (both in silicon and in design work). You shell out $500.
Which would you prefer? I'd go for option B.

Then again I'm a newb, and I could be full of crap. Haven't checked lately.

 

[Niceone]

But would those GPUs on B-model cost less than 1/4 of that from A-model? Yes more dies per wafer.. but more is needed when one card would take 4 of them. This 4 GPU model would also use more silicon.

And what about those components that are needed to connect these GPU's? For example HD3870 X2 uses PLX chip which actually is larger than RV670.
http://plaza.fi/s/f/editor/images/12_17_r680.jpg

Also PCB will also be more complex with multi GPU solutions..as will be driver developement.

 

[Silent.Sin]

But would those GPUs on B-model cost less than 1/4 of that from A-model? Yes more dies per wafer.. but more is needed when one card would take 4 of them. This 4 GPU model would also use more silicon.

And what about those components that are needed to connect these GPU's? For example HD3870 X2 uses PLX chip which actually is larger than RV670.
http://plaza.fi/s/f/editor/images/12_17_r680.jpg

Also PCB will also be more complex with multi GPU solutions..as will be driver developement.

I would actually venture a guess to say that yes, those cores OVERALL would cost less than 1/4 of the "A" die. Not only in materials cost, which is roughly 1/4 per core, but R+D and QA costs as well. Having to test out a 2 billion transistor gorilla all at once is probably not on the QA department's list of things they want to do before they die (no pun intended). I'm sure you can ask the folks behind the Itanium 2 about that. Having to test out a modular piece of silicon that is very similar to past generations (assuming an R700-B core is somewhat similar to an R600 spec core with generational improvements and changes to allow for multi-gpu connections) would be much less of a headache. After you thoroughly test an individual R700-B then all of the R700-Bs should be alright. That's ~25% of the silicon space you'd be testing versus an R700-A. You'd just have to make sure they played nice with each other after that which probably wouldn't be too huge of an undertaking as long as they've built these from the ground up to be easily connected. This doesn't even factor in the savings they'd gain from a single bad R700-A core vs a bad R700-B.

As for the connection technology I'm thinking (hoping) AMD went out of their way to develop some sort of on-die interconnect technology. HT has worked out absolutely fantastically for them. I'm sure such an idea can be translated to the GPU side of things as well. If AMD is to have any hopes of having a multi-die package work they sure as hell better have something more intelligent than slapping an entire PCI-E bridge chip near the package so the cores can talk to one another. Remember this is different than the X2. We're talking multi-core single package, not multi package. Using that chip to communicate between two packages makes sense, using it to communicate from core 1 to core 2 on the same package makes no sense. On the CPU side of things, AMD already knows this and Intel is still hurting from not having a similar technology and is stuck with the FSB/Northbridge problem, hence the gigantic caches.

Driver development is always harder with each new generation. They should be used to this by now. Just about everything these days is moving towards parallelization and developers still aren't quite comfortable with it, but you've got to admit it's getting better all the time. As for the PCB complications that really all depends on how AMD is handling the memory access and power. The main reason for adding PCB complexity lately has everything to do with memory bus width and having the paths to handle it. If they can find a smart way to feed these cores data then an overly complicated PCB might not be necessary. Power supply to each core may become a factor but that's small stuff compared to pin outs to the memory.

 

[TheBluePill]

I would actually venture a guess to say that yes, those cores OVERALL would cost less than 1/4 of the "A" die. Not only in materials cost, which is roughly 1/4 per core, but R+D and QA costs as well. Having to test out a 2 billion transistor gorilla all at once is probably not on the QA department's list of things they want to do before they die (no pun intended). I'm sure you can ask the folks behind the Itanium 2 about that. Having to test out a modular piece of silicon that is very similar to past generations (assuming an R700-B core is somewhat similar to an R600 spec core with generational improvements and changes to allow for multi-gpu connections) would be much less of a headache. After you thoroughly test an individual R700-B then all of the R700-Bs should be alright. That's ~25% of the silicon space you'd be testing versus an R700-A. You'd just have to make sure they played nice with each other after that which probably wouldn't be too huge of an undertaking as long as they've built these from the ground up to be easily connected. This doesn't even factor in the savings they'd gain from a single bad R700-A core vs a bad R700-B.

As for the connection technology I'm thinking (hoping) AMD went out of their way to develop some sort of on-die interconnect technology. HT has worked out absolutely fantastically for them. I'm sure such an idea can be translated to the GPU side of things as well. If AMD is to have any hopes of having a multi-die package work they sure as hell better have something more intelligent than slapping an entire PCI-E bridge chip near the package so the cores can talk to one another. Remember this is different than the X2. We're talking multi-core single package, not multi package. Using that chip to communicate between two packages makes sense, using it to communicate from core 1 to core 2 on the same package makes no sense. On the CPU side of things, AMD already knows this and Intel is still hurting from not having a similar technology and is stuck with the FSB/Northbridge problem, hence the gigantic caches.

Driver development is always harder with each new generation. They should be used to this by now. Just about everything these days is moving towards parallelization and developers still aren't quite comfortable with it, but you've got to admit it's getting better all the time. As for the PCB complications that really all depends on how AMD is handling the memory access and power. The main reason for adding PCB complexity lately has everything to do with memory bus width and having the paths to handle it. If they can find a smart way to feed these cores data then an overly complicated PCB might not be necessary. Power supply to each core may become a factor but that's small stuff compared to pin outs to the memory.

Couldnt have said it better myself. (really).

An on-die interconnect would probably shore up a lot of the performance loss that current Crossfire and SLI solutions have. AMD might head down that path since its easier and less expensive than producing larger and larger single GPU solutions. A great "mid/mid-high" range GPU can be put in a multi-GPU package with an excellent interconnect and come out ahead of the competition, at least for next next few product cycles. AMD is well ahead at this point and should exploit it.