AMD Bulldozer in Dec 2010?

[Voo]

Also doesn't GDDR memory need much more power and therefore also produces more heat than standard RAM? Also Intel/Amd both don't manufacture the MBs themselfes and the manufacterer's probably aren't too enthusiastic about some extra expensive cooling..

 

[cbn]

I'd think HD5570 performance would be fine. If not to cannibalize their own product lines. Graphics cards aren't going to raise the bar that much higher in 6-12 months.

The better way would be doing it Clarkdale-style, but rather than the GMCH on the other die(because its already integrated), the memory for the GPU would be on MCM within a package.

IntelUser2000,

How would this MCM'd GMCH increase memory bandwidth? You must be talking about a wider bus right? (Apologies in advance for not knowing anything about Computer Architecture)

 

[busydude]

5570 uses DDR3 memory. AMD may use a separate DDR3 slot just for the GPU core(A 1 GB DDR3 module dedicated for GPU).

EDIT: oops, didn't read IDC's post before commenting.

 

[nonameo]

5570 uses DDR3 memory. AMD may use a separate DDR3 slot just for the GPU core(A 1 GB DDR3 module dedicated for GPU).

EDIT: oops, didn't read IDC's post before commenting.

I could see 4 slots, 2 for the CPU and 2 for the GPU. Dual channel for each. It would add extra complication having 4 channels going to the CPU, but at least motherboard makers are already used to having 4 slots on there. On the other hand, after adding all of that extra cost to the motherboard, the cost of extra ram, etc..., it makes me wonder if it would be better to just buy a discrete graphics card.

 

[StrangerGuy]

I could see 4 slots, 2 for the CPU and 2 for the GPU. Dual channel for each. It would add extra complication having 4 channels going to the CPU, but at least motherboard makers are already used to having 4 slots on there. On the other hand, after adding all of that extra cost to the motherboard, the cost of extra ram, etc..., it makes me wonder if it would be better to just buy a discrete graphics card.

Quad-channel is going to be a cost nightmare to implement on a mobo, I doubt AMD are going that route unless BD is on par or superior than Nehalem.

Just look at how expensive the current triple-channel LGA1366 boards still are now.

 

[busydude]

I could see 4 slots, 2 for the CPU and 2 for the GPU. Dual channel for each. It would add extra complication having 4 channels going to the CPU, but at least motherboard makers are already used to having 4 slots on there. On the other hand, after adding all of that extra cost to the motherboard, the cost of extra ram, etc..., it makes me wonder if it would be better to just buy a discrete graphics card.

And this is not even an enthusiast chip, its a mainstream one. Also considering the cost of memory at this instance, it may be a deal breaker for a lot of folks.

 

[nonameo]

Quad-channel is going to be a cost nightmare to implement on a mobo, I doubt AMD are going that route unless BD is on par or superior than Nehalem.

Just look at how expensive the current triple-channel LGA1366 boards still are now.

Yeah I was thinking that, but if they're going to get decent performance out of this they've got to do better than the current memory sharing scheme.

 

[Idontcare]

Quad-channel is going to be a cost nightmare to implement on a mobo, I doubt AMD are going that route unless BD is on par or superior than Nehalem.

Just look at how expensive the current triple-channel LGA1366 boards still are now.

The memory arrangement implemented for Llano need be no more expensive or complicated than that currently involved in existing discrete graphics cards of similar price segment.

Also be careful about conflating price of a product with apparent feature-set of the product...LGA1366 is expensive in part because the market size is tiny, and was always envisioned to be tiny, so mobo makers priced the boards appropriately to recoup their fixed development costs. To be sure triple-channel boards cost more to design and produce than dual-channel, but not $100 more...

 

[Terzo]

How are bulldozer and llano related? They aren't the same chip, are they?

 

[nonameo]

How are bulldozer and llano related? They aren't the same chip, are they?

llano is, from my understanding, based on phenom II, but bulldozer will be a new architecture with no on die graphics.

 

[busydude]

How are bulldozer and llano related? They aren't the same chip, are they?

No, they are not the same chip. Llano is an improved Athlon II architecture with a GPU on die. It is mainly for laptops and mainstream desktops. Whereas, bulldozer is a completely new architecture with up to 8 integer cores and no GPU on board. Both are 32 nm SOI parts with power gating technology similar to the one Intel using currently.

 

[Terzo]

llano is, from my understanding, based on phenom II, but bulldozer will be a new architecture with no on die graphics.

No, they are not the same chip. Llano is an improved Athlon II architecture with a GPU on die. It is mainly for laptops and mainstream desktops. Whereas, bulldozer is a completely new architecture with up to 8 integer cores and no GPU on board. Both are 32 nm SOI parts with power gating technology similar to the one Intel using currently.

Ah, thanks. I guess this is starting to thread derail, but what would be the intel counterparts? I'd imagine bulldozer would be pitted against sandy bridge, except it lacks an on board gpu.

 

[nonameo]

Ah, thanks. I guess this is starting to thread derail, but what would be the intel counterparts? I'd imagine bulldozer would be pitted against sandy bridge, except it lacks an on board gpu.

Ah yeah, I slipped... it would be based on athlon II and not phenom II because I do believe it will not have L3

 

[IntelUser2000]

IntelUser2000,

How would this MCM'd GMCH increase memory bandwidth? You must be talking about a wider bus right? (Apologies in advance for not knowing anything about Computer Architecture)

Well, yea. They can probably use Hypertransport link to connect to the MCM memory which will save the trouble of having to create extra traces on the motherboard, and the problem of finding vendors who will make the GDDR module. Rather than having ~25GB/s shared, maybe they can do 2x dedicated.

 

[alyarb]

Llano's architecture brings a number of interesting peculiarities because of its high level of integration but none of these peculiarities are performance opportunities. You can do whatever you want with the memory bus, even if it increases the cost of a traditional shared DDR3 board by $60-75, and you will still be shader limited, still be slower than a 5670 or the 5700's***, and you will have a 60% more expensive product that's only 20% faster than with the shared RAM.

When you realize that people who use integrated platforms very very rarely play games, the added cost simply doesn't make sense because none of the performance increases will be noticed or appreciated. The most you should hope for from Llano is an accelerated sideport. That should give everybody a nice 10-15% for hybrid crossfire, especially if there's an elegant way of sharing the discrete VRAM among the Llano and Radeon shaders.


*** unknown Llano clock domains. i'm thinking 3 GHz core, 2.6 GHz uncore, and at least 1 GHz shaders.

 

[VirtualLarry]

*** unknown Llano clock domains. i'm thinking 3 GHz core, 2.6 GHz uncore, and at least 1 GHz shaders.

What are the odds that the shaders would run at the CPU core clock? I realize not likely, but it is fun to speculate. That could prove to be a lot more powerful than one would think, if true.

 

[cbn]

What are the odds that the shaders would run at the CPU core clock? I realize not likely, but it is fun to speculate. That could prove to be a lot more powerful than one would think, if true.

480 stream processors @ 3000 MHz? That would be faster than HD5870.

 

[alyarb]

I'm going to give them the benefit of the doubt and assume that they are intimately familiar with their product. they have years of data that tells them how much bandwidth their shaders need at a given frequency. they should know exactly where the happy medium exists in the power/frequency relationship and they are targetting an economical, efficient compromise that can run reliably underneath the worlds worst heatsink. they are not trying to optimize llano for graphics performance. entry level.

 

[cbn]

they are not trying to optimize llano for graphics performance. entry level.

That would be disappointing considering the lack of general use/productivity software for ATI GPUs (at this time).

P.S. According to this article AMD/ATI plan on doing away with discrete graphics cards. This makes me think they have plans for optimized graphics performance.

 

[yh125d]

What are the odds that the shaders would run at the CPU core clock? I realize not likely, but it is fun to speculate. That could prove to be a lot more powerful than one would think, if true.

Nil. Theres a reason modern shaders top out around 1gHz, and it's not cause we aren't feeding them enough voltage. The architecture of a shader unit just doesn't lend itself to high frequencies

 

[alyarb]

That would be disappointing considering the lack of general use/productivity software for ATI GPUs (at this time).

P.S. According to this article AMD/ATI plan on doing away with discrete graphics cards. This makes me think they have plans for optimized graphics performance.

llano is a necessary first step to leverage their shader-on-x86 strategy. it's not gpgpu, it's heterogeneous. integrated graphics have never been fast and that isn't going to change with this generation. making it fast requires all the extra crap you guys are all speculating about, and it's just silly. no one is MCMing anything on this chip. they're doing it because it gets their foot in the heterogeneous door ahead of intel and because it costs less than making IGPs at TSMC. it's business they can throw at themselves vis a vis GF and motherboard margins should widen a tad as well. even if the laws of physics permitted it, a 3ghz graphics powerhouse would totally cannibalize the lower discrete segments.

I want to see some kind of hardware SSE-to-radeon compiler. that is what heterogeneous means to me.

 

[cbn]

I want to see some kind of hardware SSE-to-radeon compiler. that is what heterogeneous means to me.

If AMD has this compiler ready at the time of Llano's lauch, what kind of programs (besides games) could be run on the APU?

We are about talking about speeding up the processing time of everyday programs meant to run strictly on CPUs right? Would this include all programs? Or are we talking about only certain modern programs?

Thanks.

 

[cbn]

http://www.tomshardware.com/reviews/athlon-ii-x2-260-athlon-ii-x3-445,2629-3.html

Here is the link where HD5570 discrete performance was mentioned for Llano.

"Information from AMD suggests that llano's integrated graphics core may perform on par with the discrete Radeon HD 5570. This is very, very powerful for an integrated part and may truly bring 1680x1050 gaming to the masses."

 

[nyker96]

http://www.tomshardware.com/reviews/athlon-ii-x2-260-athlon-ii-x3-445,2629-3.html

Here is the link where HD5570 discrete performance was mentioned for Llano.

"Information from AMD suggests that llano's integrated graphics core may perform on par with the discrete Radeon HD 5570. This is very, very powerful for an integrated part and may truly bring 1680x1050 gaming to the masses."

I find this report a bit hard to believe, considering most of integrated solutions on motherboards are just soso. To put such a powerful gpu core into amd's first fusion product seems to be quite a leap forward. I hope it is true that they are able to do this and keep everything under a reasonable power and heat envelop.

 

[alyarb]

heh llano is going to be way faster than 785G. don't get me wrong. but you guys are speculating to an extreme.


a fusion compiler is still sci-fi as far as I know, but the concept should be on everyones radar.