AMD "Fusion" - My Opinions - CPUs

I think the project "Fusion" is for the mobile space.

Integrating CPU, GPU and memory controller into one piece of silicon with power-saving mechanisms will help AMD to gain more mobile market share.
I have read a news for one of the reason for Turions' higher power consumption numbers - the integrated memory controller.

There is absolutely not need to fuse CPU and GPU together in a desktop for saving space and power. But it is critical for notebooks to have a small size.

Add a reply

Register or log in to remove.

With the "fusion" some architectural parts like the FPU will be shared and more often and more effectively used. Also there would be less buses(no AGP/PCIe graphics bus and VRAM lanes), the ODMC will be used for both CPU & GPU data transfers.

Reply to gOJDO

Quote :

With the "fusion" some architectural parts like the FPU will be shared and more often and more effectively used. Also there would be less buses(no AGP/PCIe graphics bus and VRAM lanes), the ODMC will be used for both CPU & GPU data transfers.

But the major problem is the die size 8O

Reply to qcmadness

I have to say I'd agree with you. You don't really hear alot of people saying "i need to upgrade my laptop's video card" since its a pain in the neck.

I can see a possible desktop application, but I think there's going to be too much resistance because of the higher cost of the CPU/GPU and that everyone will stick to the easily upgradeable discrete video cards.

The extra floating point power would be really good for the scientific community though.

Reply to nobly

Quote :

I have to say I'd agree with you. You don't really hear alot of people saying "i need to upgrade my laptop's video card" since its a pain in the neck.

I can see a possible desktop application, but I think there's going to be too much resistance because of the higher cost of the CPU/GPU and that everyone will stick to the easily upgradeable discrete video cards.

I think your missing the business market. 90-95% of business PC's do not need a discrete card. I can tell you that having the graphics on the CPU or the MoBo is pretty much the same for that market, whatever works and is cheapest FTW.

For users who needs graphics cards the situation is different, but considering what most people do computing wise its a relatively minor segment of the user base in terms of overall numbers. I mean come on if Intel is the #1 graphics company obviously there's a large market that doesn't need discrete cards.

Reply to kamel5547

Very true, I did miss that.

But why integrate it on-chip? Onboard graphics are already there and these office users don't really need that touted extra power to get the GPU onto the chip die. If they have onboard and are happy, the only reason they'd choose a GPU/CPU chip is cost. I'm sure AMD and ATI could chug out an onboard graphics solution alot quicker than an on-chip solution.

And I think that by putting the CPU/GPU onto a single die will be much more expensive than the on-board graphics solution. Each wafer would yield less CPU/GPU chips since the die size would be bigger, right?

Reply to nobly

Moving GPU to CPU would probably be closer to a new architecture than just sticking two separate systems on the same silicon. Now, in addition to x86 ALU, x86 FPU, and SSE, the chip will have additional xx vertex units.

Remember back when the CPU did the vector math and the GPU just created pixels? The 3D math needed for the amazing graphics of the VooDoo era cards was not efficient on x86 processors, but getting a completely new processor design off the ground is impossible, witness Sparc, Itanium, etc. Hence the GPU was born.

There are also a several other applications that can take advantage of this type of processing. The supercomputers with Opterons and Cell processors paired is an example. Also, certain scientific software is being written to run on GPU's, like folding at home. And lets not forget physics engines. The 'GPU in the CPU' could be used for physics while the high dollar GPU does graphics.

Just as once expensive options like math coprocessors were integrated into the CPU, so it happens again with vertex units.

Reply to mr_fnord

Couple points here, first off, the IMC is really not that big of a factor in power consumption, the Intel system has a northbridge with a memory controller, too, lol. Second, most people with desktops wouldn't mind having the ability to run IE and Word without a graphics card, even if they'll be using a "real" GPU for most applications.

Reply to shabodah

Quote :

Couple points here, first off, the IMC is really not that big of a factor in power consumption, the Intel system has a northbridge with a memory controller, too, lol. Second, most people with desktops wouldn't mind having the ability to run IE and Word without a graphics card, even if they'll be using a "real" GPU for most applications.

The problem for IMC is that when a data is fetched from the main memory, the CPU is involved. But with a northbridge, the data-fetching will by-pass CPU .

Reply to qcmadness

The total die size would probably be slightly smaller than the combined die sizes of the two chips. There would be some duplicated features that could be removed. Yields can be size dependent though, so one larger chip may have lower yields than two smaller chips.

In the laptop realm this would lead to some power savings, plus there would be fewer busses and subcomponents to electrify. Lower cost graphics also aren't necessarily made using state fo the art fabs, so a 45nm integrated GPU and CPU would have power savings and greater performance than a 65nm GPU.

Reply to mr_fnord

I dont think the die will need to get much bigger, in fact if they do it on the 45nm process they could get a dual core cpu and a gpu in a 3 core solution. in any case the idea is to get them working together as well so they both become more powerful, when you need graphic power the cpu will help out, when you need cpu the gpu will help. mabey they'll do a two core, one cpu, one gpu and set up a power load/process share between them. there are a lot of possibilities and ideas they could impliment doing this, and whatever they do will most definately use less W overall than having a cpu and gpu seperate.

Reply to jrabbitb

On a laptop with far fewer components than an enthusiasts desktop, what exactly is going to access the memory often that is not the CPU itself? And on the other end of the spectrum, With Intel's bus, when the CPU needs to access memory, but it and the northbridge would have to be "awake."

Reply to shabodah

Quote :

On a laptop with far fewer components than an enthusiasts desktop, what exactly is going to access the memory often that is not the CPU itself? And on the other end of the spectrum, With Intel's bus, when the CPU needs to access memory, but it and the northbridge would have to be "awake."

But a northbridge consumes less power than a CPU, definitely.

Reply to qcmadness

A northbridge consumes less power than a fully alive CPU, certainly, but in these notebooks and the different "sleep" states the CPU's have, things the ability to work in abnormal ways. Heck, I've seen Intel north and south bridges consuming 25-30 watts. But as I asked before, WHAT exactly is accessing the system memory that is NOT the CPU?

Reply to shabodah

[quote=qcmadness]But a northbridge consumes less power than a CPU, definitely.[/quote]

Did intel systems stop shipping with CPUs? Does this magic Northbridge do everything now? The only hope of a power savings from a non-integrated memory controller is with AGP or direct memory access. Even then, AMD's IMC does not create a massive workload for the CPU. Just because the IMC is operating doesn't mean the CPU is at full load.

There may be advantages to having a non-integrated MC, ease of upgrading memory, lower power consumption in the CPU package, smaller CPU die size for higher yields, but it is unlikely that the IMC consumes that much more power than a Northbridge.

Reply to mr_fnord

Quote :

A northbridge consumes less power than a fully alive CPU, certainly, but in these notebooks and the different "sleep" states the CPU's have, things the ability to work in abnormal ways. Heck, I've seen Intel north and south bridges consuming 25-30 watts. But as I asked before, WHAT exactly is accessing the system memory that is NOT the CPU?

Integrated graphics, harddisks, etc...

Reply to qcmadness

How often is a DMA device active when the CPU is asleep? Does the HD get bored and start randomly loading data into a DMA block for fun?

Reply to mr_fnord

Integrated graphics are accessing the system memory while the CPU is in a sleep state?!?

Reply to shabodah

Quote :

The total die size would probably be slightly smaller than the combined die sizes of the two chips. There would be some duplicated features that could be removed. Yields can be size dependent though, so one larger chip may have lower yields than two smaller chips.

Exactly my point. If its a bigger die size, then 1) you get less products because your wafer is probably a constant size and 2) you may have worse yields. Plus putting the GPU and CPU on one chip makes logic more complex. At 45nm it might be okay, but they have to get to 45nm first...

Quote :

mabey they'll do a two core, one cpu, one gpu and set up a power load/process share between them. there are a lot of possibilities and ideas they could impliment doing this, and whatever they do will most definately use less W overall than having a cpu and gpu seperate.

Yeah this is what I think will probably happen once everything starts going multi-core. A few dedicated floating point cores, a few dedicated int computing cores, etc. Its not terribly new, the cell processor is loosely based on this, but with more floating point cores since its graphics intense.

Quote :

A northbridge consumes less power than a fully alive CPU, certainly, but in these notebooks and the different "sleep" states the CPU's have, things the ability to work in abnormal ways. Heck, I've seen Intel north and south bridges consuming 25-30 watts. But as I asked before, WHAT exactly is accessing the system memory that is NOT the CPU?

Can you provide a link for that? I'm interested to see how they tested it to exactly find the power consumption on the north bridge. THanks!

Reply to nobly

No I don't have links. But this was for a laptop system with onboard graphics, too, so my point was just that a northbridge uses electricity, too. It's not some magic device that runs off "happy thoughts."

Reply to shabodah

Quote :

Can you provide a link for that? I'm interested to see how they tested it to exactly find the power consumption on the north bridge. THanks!

You can find more about Intel chipsets in
www.intel.com
:wink:

Reply to qcmadness

Quote :

Good opinion, I will add my 2 cents....

Fusion is a good idea, and will find a place in the market.... it has two real applications, low end-entry level chips where graphics/CPU combo can reduce costs overall and in high speed FPU related work where the parallel capabilities of the GPU can be exploited (at least this what I have gathered from reading several articles on the topic).

AMD fans who expect a Radeon R600 type GPU married to a CPU and give great gaming performance though will be rudely dissapointed. At 80 nm, current GPU die sizes are in excess of 250 and 300 mm^2, goint to even 45 nm puts the GPU portion of a die at 100 to 150 mm^2, it would be cost prohibitive to take this approach. Hence, AMD's announcement that discrete GPU solutions will not be going away anytime soon.

I think the GPU-on-CPU will be the same class as Crush-51...

Reply to qcmadness

First off we were talking about laptop and budget oriented, basically barebones, desktops. In these applications it is unlikely that you'd need to access system memory when the cpu is not working. I have no illusions about AMD having a northbridge in their system. Just pointing out that if the CPU and chipset are both "fully awake," it matters not where the memory controller is, you'll be using the same amount of energy (give or take depending on specifics).

Reply to shabodah

Quote :

First off we were talking about laptop and budget oriented, basically barebones, desktops. In these applications it is unlikely that you'd need to access system memory when the cpu is not working. I have no illusions about AMD having a northbridge in their system. Just pointing out that if the CPU and chipset are both "fully awake," it matters not where the memory controller is, you'll be using the same amount of energy (give or take depending on specifics).

DMA can be done when CPU is IDLE.
http://en.wikipedia.org/wiki/Direct_memory_access

Reply to qcmadness

I think AMD fusion is a great laptop and lower graphic desktops. Also this wont be out until AMD gets to 45nm so die size wont be as much of a problem. I cant wait to see what these things are like. I have heard a few things about this and it does seem to be more for the laptop, however there are some desktop and server applications for this processor.

Reply to Frozen_Fallout

Quote :

Good opinion, I will add my 2 cents....

Fusion is a good idea, and will find a place in the market.... it has two real applications, low end-entry level chips where graphics/CPU combo can reduce costs overall and in high speed FPU related work where the parallel capabilities of the GPU can be exploited (at least this what I have gathered from reading several articles on the topic).

AMD fans who expect a Radeon R600 type GPU married to a CPU and give great gaming performance though will be rudely dissapointed. At 80 nm, current GPU die sizes are in excess of 250 and 300 mm^2, goint to even 45 nm puts the GPU portion of a die at 100 to 150 mm^2, it would be cost prohibitive to take this approach. Hence, AMD's announcement that discrete GPU solutions will not be going away anytime soon.

Hi Jack, good to see in this one.
Just a few thoughts:
I can totally understand the entry-level point.
But since a discrete card will be more powerful (in terms of computational power), wouldn't it make more sense for people to just have programs that access the discrete card's floating point capabilities? Or would it be enough to have that lower-performance capabilities in the hypertransport loop for quick access (i.e. the lower performance capability would be given a big enough edge on-die to offset the higher performance discrete card)? (Assuming they're going to use hypertransport, of course).

Granted, it might be good to just have both for an increase of floating point power.

I'm convinced its a good idea, but I wonder how the market will embrace it and if it'll end up becoming a niche.

I have to wonder if it'll be cheaper to produce than on-board graphics - if not, the entry level market may not be a very good target.

Reply to nobly

The fewer the chips, the smaller the PCB, lol. I think this type of tech will be what helps bring us to smaller and smaller computers. IIRC, Intel is still the largest reseller of GPU's, even if they are IGP and do "suck alot." I see that as a huge market. The average person is not on here posting, and is not using a descrete GPU.

Reply to shabodah

I really don't think that AAMTDI would undertake a plan to double the die size for a combined product with no benefit other than a single chip in place of two. There have to be advantages in production cost and performance. Now, if you took an RV380, stripped off the PCIe interface and whatever other extraneous silicon you could and built it into a CPU you'd still have 8 pixel pipelines capable of decent 3d graphics and possibly more general calculation work.

Discrete components generally cost more and use more power, all of that additional packaging, busses, etc. That said, your integrated solution needs to be competition-proof, or in 6 months you'll be creating a product that wastes xx% of its silicon on unused features.

For tasks like physics calculations you also have the chicken and egg problem. Without a large installed base, no company will design software to drive the hardware. GPUs were the same way before directX. Without known functionality to develop to, no software company would write software hoping that the hardware becomes common. If this GPU on CPU has a greater reach than strictly GPU functions, and AMD gets a fairly large installed base, and software starts being written for it, this would put intel in a follow the leader position just like MMX did to AMD and the K6.

Reply to mr_fnord

The other thing about this is being able to drop the cost of the GPU by integrating the CPU and GPU. Also from what I hear I think they get some small benefits form putting the two together I cant remember at this time what they are all together but I think tom has something on all this. Anyway I think this is a great idea for AMD to pick up the mid-market by having a good GPU CPU combo. Most people 60%+ (yes I'm pulling this number out of my ass but you get the idea) don't care about having the best of the best in there computer for the GPU and just want something that can play some games and use the Internet and play moves small stuff like that. And when it comes to the laptops this makes it allot easier to put a good gpu in and still have a small case since it isn't a card and if they can have descent GPUs in the laptops that are above the normal laptops this would be a great selling point.

Reply to Frozen_Fallout

AMD "Fusion" - My Opinions

Forum CPU & Components : CPUs - AMD "Fusion" - My Opinions

Please mind

Focus On

AMD "Fusion" - My Opinions

Forum CPU & Components : CPUs - AMD "Fusion" - My Opinions

Please mind