Is AMD fusion going to be a legit gaming platform?

[BD2003]

Hearing a lot of rumbling about AMD merging the CPU and the GPU. It seems to me like intel is on the same path with their current chips and upcoming sandy bridge.

I have little faith that intel is going to integrate decent graphics into their chips. But AMD seems like they might be a different story, and even though theyre supposedly dropping in Q4, I cant seem to find any reliable info on what theyre targeting fusion at.

Are they attempting to fuse gaming-level graphics onto the CPU, or should I be expecting something more along the lines of typical underpowered integrated graphics? Basically, or they doing something new here, or is this the same old integrated by another name?

 

[LoneNinja]

It's been discussed here before, and the rumors lead to a Redwood level gpu integrated with a 32nm variant of Athlon II X4. If you don't know what redwood is, it's the gpu used for the radeon 5570/5670. Several times stronger than current IGP solutions, and good enough for most games, but true gamers will still want to get discrete cards. Realistically it'll remain that way for a long time, as the pure size and power requirements of our high end cards won't be integrated, not to mention system memory allocation rather than dedicated Vram.

 

[Rifter]

I dont see it doing well for gaming. I do see it doing very well in the HTPC/Casual non-gaming market though.

I do think it will be better than current IGP's, i can even see it being 2-3 times more powerful or a bit more. However i think the extra heat and power from the combined GPU and CPU will make it difficult to OC compared to a discreet GPU setup and therefor hold back performance.

And a 5570 is by no means a gaming card, and thats assuming they even hit 5570 performance.

At the end of the day though we will just have to wait and see.

 

[jvroig]

For gamers satisfied with console graphics and performance, I'd say Fusion will most likely deliver enough performance (if the 5670 rumors turn out accurate)

 

[BD2003]

So if it seems likely that graphics performance isnt going to be competitive, what is so special about the merging of the CPU and GPU that we cant simply accomplish with a good CPU and a discrete card?

I've read a few articles that seem to imply that this is going to be something fundamentally different and special, that merging the GPU onto the die is going to produce something more than just your typical integrated graphics. Is there something I'm missing?

 

[cbn]

I've read a few articles that seem to imply that this is going to be something fundamentally different and special, that merging the GPU onto the die is going to produce something more than just your typical integrated graphics. Is there something I'm missing?

It might be fusion is better at performing GPGPU tasks if the iGP is able to share some of the CPU cache.

 

[jvroig]

So if it seems likely that graphics performance isnt going to be competitive, what is so special about the merging of the CPU and GPU that we cant simply accomplish with a good CPU and a discrete card?

I suppose if you are concerned about gaming performance, then nothing special will happen at all. They are envisioning Fusion to be within the lines of "do tasks the CPU is good at on the CPU, while tasks that are better suited for a GPU on the GPU (which is now a part of the CPU)". If I remember correctly, Anand also referenced this, when he talked about the shared FPU between 2 Bulldozer Int cores - he said that FPU bandwidth for now is ample, and eventually it will be offloaded to a fusion GPU anyway which would be more efficient at it. (Paraphrased, just typing it from memory).

The real value for Fusion then is not in what we can do better from the things we already do now, but it may allow for things that are better in the future, when FPU performance skyrockets within CPUs due to Fusion-type integration. When each CPU acts like a decent GPU as well, GPGPU computing won't be limited to having a discrete card and its target audience/base will expand, which will affect development of software for it (so all apps that are still performance-starved and are limited to using only the CPU may benefit from Fusion)

In that way, it's not a revolution, but it's one step that can result in a lot of great things in the future.

Since there is no Fusion product on hand right now, I cannot say how promising or how great this step actually is, or if it can actually make all those visions possible at all. For all we know, it can turn out to be nothing more than another hype machine on steriods, offering nothing more than just "IGP on the CPU instead of the mobo". At least, even that would mean lesser crap on the mobo.

 

[MagnusTheBrewer]

It's an entirely new architecture and folks are whining about will it be a gaming platform? Why don't we give it a few years to get the kinks and future possibilities worked out first.

 

[OCGuy]

No....no.

AMD will even tell you....no.

They didnt go out and spend a bunch of $$ acquiring a discreet GPU company if they thought they could do it themselves.

 

[BD2003]

I suppose if you are concerned about gaming performance, then nothing special will happen at all. They are envisioning Fusion to be within the lines of "do tasks the CPU is good at on the CPU, while tasks that are better suited for a GPU on the GPU (which is now a part of the CPU)". If I remember correctly, Anand also referenced this, when he talked about the shared FPU between 2 Bulldozer Int cores - he said that FPU bandwidth for now is ample, and eventually it will be offloaded to a fusion GPU anyway which would be more efficient at it. (Paraphrased, just typing it from memory).

The real value for Fusion then is not in what we can do better from the things we already do now, but it may allow for things that are better in the future, when FPU performance skyrockets within CPUs due to Fusion-type integration. When each CPU acts like a decent GPU as well, GPGPU computing won't be limited to having a discrete card and its target audience/base will expand, which will affect development of software for it (so all apps that are still performance-starved and are limited to using only the CPU may benefit from Fusion)

In that way, it's not a revolution, but it's one step that can result in a lot of great things in the future.

Since there is no Fusion product on hand right now, I cannot say how promising or how great this step actually is, or if it can actually make all those visions possible at all. For all we know, it can turn out to be nothing more than another hype machine on steriods, offering nothing more than just "IGP on the CPU instead of the mobo". At least, even that would mean lesser crap on the mobo.

Hmm...so if I'm understanding this correctly, theyre quite literally fusing the CPU and GPU? This isnt just putting a separate CPU and GPU on the same die and shortening the distance between them and sharing cache, but taking aspects of each and making a hybrid chip that cant be accurately described as just another CPU?

Will programs need to be coded specifically to take advantage of or even work on this? Or is there the potential for GPGPU benefits without having to code specifically for it...that the CPU/GPU/APU/whatever they call it just runs appropriate instructions on whatever part of the architecture is best for the job?

 

[busydude]

Few months back, at COMPUTEX '10, AMD has shown a live demonstration of Ontario(A low power APU meant for netbooks). They showed a small gameplay clip of the AvP Dx11 game.

 

[IntelUser2000]

So if it seems likely that graphics performance isnt going to be competitive, what is so special about the merging of the CPU and GPU that we cant simply accomplish with a good CPU and a discrete card?

I've read a few articles that seem to imply that this is going to be something fundamentally different and special, that merging the GPU onto the die is going to produce something more than just your typical integrated graphics. Is there something I'm missing?

It's just about providing more value for the CPU. The benefits of multi-cores are becoming lower and lower as the amount of cores increase. Integrating a GPU that can execute those tasks well while sharing resources allow a much more efficient combination than similar, discrete setup.

There's not that many better alternative on laptops, where the GPUs are currently bound by TDP. Llano and Sandy Bridge setups will offer mid-range discrete graphics performance at a price that only 3-year old value discrete graphics variants can touch, and with power usage low enough to allow some gaming on battery.

 

[superccs]

You guys remember 3dfx? They would sell you a 3d addon card for running games. It think with a fusion CPU/gpu switchable graphics might be able too make some progress. Having a hulking 100-200w discrede gpu running idle is still wasteful ... if you can run Eaton from CPU/gpu. And completely shut down your gpu that would be cool.

 

[ModestGamer]

This is ultimately aimed at chipsets and OEM builders. Reducing parts count saves $$$$ and it they get acceptable performance doing that then this is how it is going to go from here forward. The real limitation is socket, TDP and process. If they get down to say 16nm maybe we will see current generation 5970 quality or more powerful GPU integrated into the CPU. Hell for all we know they make go beyond the added core logic and actually build a massive G/CPU in a much more integrated fashion. This is a strat but the OEM system builder are clamoring for high margins and cheaper boards and lessparts are the way to get there.

 

[sandorski]

Eventually, probably yes.

 

[cbn]

This is ultimately aimed at chipsets and OEM builders. Reducing parts count saves $$$$ and it they get acceptable performance doing that then this is how it is going to go from here forward. The real limitation is socket, TDP and process. If they get down to say 16nm maybe we will see current generation 5970 quality or more powerful GPU integrated into the CPU. Hell for all we know they make go beyond the added core logic and actually build a massive G/CPU in a much more integrated fashion. This is a strat but the OEM system builder are clamoring for high margins and cheaper boards and lessparts are the way to get there.

No doubt cheap hardware helps system builder bottom lines, but how far can that really go when MS Windows is still a fixed cost?

 

[ModestGamer]

No doubt cheap hardware helps system builder bottom lines, but how far can that really go when MS Windows is still a fixed cost?

how many components can you remove from a MB if your building net appliances or consumer gear? Thats where the market is and has been heading for sometime with the home PC market. If the kids can play scoobys great adventure and mom can check her email and play farmland apps. Thats all it needs to do and do it well.

 

[cbn]

how many components can you remove from a MB if your building net appliances or consumer gear? Thats where the market is and has been heading for sometime with the home PC market. If the kids can play scoobys great adventure and mom can check her email and play farmland apps. Thats all it needs to do and do it well.

It sounds like you are describing an ARM phone based OS device?

If so, wouldn't the costs of building such a system be impacted more by hardware costs rather than MS Windows costs? (re: Android, for example, is essentially a free install).

 

[Scali]

I have little faith that intel is going to integrate decent graphics into their chips.

Actually, Intel did some demonstrations with early Sandy Bridge chips, and the graphics were pretty good for Intel's standards.
They may just be able to compete with the ultra low-end discrete card market (stuff like the Radeon 5400 or so). On which you may just be able to play some games, as long as you turn resolutions and detail down far enough

Are they attempting to fuse gaming-level graphics onto the CPU, or should I be expecting something more along the lines of typical underpowered integrated graphics?

What they're essentially doing, is bundling a Radeon 5570-ish GPU onto the CPU (which will probably be slower than a discrete version since it needs to share the memory and controller).
So yes it's quite a step up from conventional IGPs... but no, I personally don't really see a 5570 as a good gaming GPU. If I look at these benchmarks:
http://www.anandtech.com/show/2935/5
It would appear that you still need to turn down the resolution and detail considerably to reach framerates of 30-50 fps.

 

[IntelUser2000]

Actually, Intel did some demonstrations with early Sandy Bridge chips, and the graphics were pretty good for Intel's standards.
They may just be able to compete with the ultra low-end discrete card market (stuff like the Radeon 5400 or so). On which you may just be able to play some games, as long as you turn resolutions and detail down far enough

I liked the World of Warcraft demo. Current HD Graphics are in average 2x faster than the GMA 4500, but in WoW its only 30-40% faster. On the WoW demo Intel has done, the GMA 4500 was getting 2-3 fps with everything up at max while the Sandy Bridge based laptop wasn't getting any noticeable lag. On a laptop too.

 

[Scali]

Hmm...so if I'm understanding this correctly, theyre quite literally fusing the CPU and GPU? This isnt just putting a separate CPU and GPU on the same die and shortening the distance between them and sharing cache, but taking aspects of each and making a hybrid chip that cant be accurately described as just another CPU?

No, they're not fusing CPU and GPU, at least not on the short term.
It's quite literally a Radeon GPU copy-pasted onto an Athlon II die.

There is not much more you can do on short notice anyway, since any closer integration would require both to share the same instructionset. So the GPU would become some kind of SSE-extension to the CPU.
Anything else would be nothing different from a second processor on a separate bus, with its own memory pool for instructions. So it might as well be a discrete card.

 

[taltamir]

Is AMD fusion going to be a legit gaming platform?

No, fusion is going to provide cheaper, lighter, smaller, and longer battery life laptops.
it will provide cheaper low end desktops.

it has nothing to do with gaming and will provide absolutely nothing for a gamer. (actually, for a gamer / high end computer it will be a DRAWBACK and provide a weaker computer overall)

fusion is merely next gen integrated graphics. (same goes for the intel version, which, btw, is arriving earlier than AMDs)

 

[heyheybooboo]

It might be fusion is better at performing GPGPU tasks if the iGP is able to share some of the CPU cache.

It looks as if the GPU (at least on Liano) will have its own buffer and HT link to the memory controller.

Hmm...so if I'm understanding this correctly, theyre quite literally fusing the CPU and GPU? This isnt just putting a separate CPU and GPU on the same die and shortening the distance between them and sharing cache, but taking aspects of each and making a hybrid chip that cant be accurately described as just another CPU?

Will programs need to be coded specifically to take advantage of or even work on this? Or is there the potential for GPGPU benefits without having to code specifically for it...that the CPU/GPU/APU/whatever they call it just runs appropriate instructions on whatever part of the architecture is best for the job?

The GPU die will be independent of the the CPU.

You are correct in that the new AVX instruction set is the first step in taking advantage of the compute power BUT the question becomes how well (efficiently) can the CPU hand off the process(es) to the GPU.

Probably not so great with Liano (Stars cores with Redwood GPU) as compared to the 'true' 2011 Bulldozer.




--

 

[Scali]

The GPU die will be independent of the the CPU.

In AMD's case, GPU and CPU will be on the same die.

 

[heyheybooboo]

In AMD's case, GPU and CPU will be on the same die.

Sorry ... I meant core as the poster was referencing 'fusing' the cpu and gpu.




--