How many years till we see APUs as standalone gaming PCs?

[alyarb]

the IGP is still a radeon. it will not run at the same clock as the Llano CPU; not even half the clock.

The K10 cores are 35 million transistors each (not including L2), but the Llano device itself is close to a billion transistors total. Estimating the die area is not as simple as adding the 5670 and Propus areas together.

 

[grimpr]

Some quotes of ATI's mr.Catalyst from an interview at pcinpact.

PCi : MSI recently announced a motherboard with Lucid Hydra Chip, allowing Multi-GPU between two Radeons, Radeon and GeForce... without using Catalyst CrossFireX technology. What are AMD's thoughts about it ?

T.M. : Interesting, but I don’t know enough about it to comment – I actually haven’t ever seen it. Our new ATI CrossFireX architecture is now capable of mixing any combination of ATI Radeon (or AMD IGP) graphics accelerators, and developing new ways to break up the workload.

OpenCL / GPU Computing
OpenCL support is included in Catalyst, but users still need to use the SDK to enable it. Why ? When will it be different ?

T.M. : We don’t put all our available software in our general Catalyst package. In that package is only what is needed to get your GPU working (display driver and CCC).

It is an optional download for the developers or users that need it. We have millions of people downloading Catalyst every month, I think its unfair to make the package size larger and put in stuff that a lot of people may not need.

In the future when a lot of OpenCL applications exist then we may revisit this decision.

PCi : Aren't you afraid that developers look after other GPU Computing solutions for their software ? (DirectCompute, CUDA...).

T.M. : Not at all.

We are certain developers will choose DirectCompute and OpenCL since they want to support their software on the full PC ecosystem install base. It is simple economics really.

If you owned a software company would you develop a solution that works on only 30% of PC’s in the world ?

PCi : NVIDIA is planning to provide « C++ like » code support for Fermi, and there is a lot of wrappers available for C for CUDA. Do you plan the same thing for Stream ?

T.M. : I can’t get into unannounced product or technology, but we are looking at many things for ATI Stream and OpenCL support.

PCi : In the Press PDF, you mentioned that there will be a new, separated, multi-GPU driver. What's the aim of this ?

T.M. : Fusion. The new re-architecture of Crossfire that we started is all about being able to use our APU with our GPU’s in the future. It will be really cool ;-)

http://www.pcinpact.com/dossiers/amd-radeon-catalyst-interview-terry-makedon/167-1.htm

 

[cbn]

T.M. : Fusion. The new re-architecture of Crossfire that we started is all about being able to use our APU with our GPU’s in the future. It will be really cool ;-)

http://www.pcinpact.com/dossiers/amd-radeon-catalyst-interview-terry-makedon/167-1.htm

This technology and "sideport" for synchronizing GPUs sounds very interesting.

Being able to link 480 IGP stream processors with a midrange HD6xxx chip will no doubt interest many buyers. Assuming HD6770 is the equivalent of 1600 stream processors we would be talking a bump to 2080 stream processors with the addition of the IGP. That is a 30% boost.

 

[hans007]

the new 40 shader 880G (and really also the 780 / 785G) supposedly supports linking to the 80 shader 5450 so maybe its all in software...

 

[deimos3428]

T.M. : Interesting, but I don’t know enough about it to comment – I actually haven’t ever seen it. Our new ATI CrossFireX architecture is now capable of mixing any combination of ATI Radeon (or AMD IGP) graphics accelerators, and developing new ways to break up the workload.

Is this accurate? Will we be able to CF 48xx and 5xxx cards, for example? (If so, the 5xxx cards just got a lot more interesting, as they've gone from "replacement single card" to "additional CF card".)

 

[lifeblood]

I think you expecting to much from Llano in the area of graphics. Initially I would expect it to work only a little better than an IGP. So while the cost of the CPU will go up, the cost of the chipset will go down. Because the GPU is on die with the CPU, graphics performance will increase but not incredibly.

Llano will probably be a better CPU because it now has a far more power floating point unit (the GPU) built in to it, but it probably won't be stellar as a GPU. This thing really is designed for business PC's, not gaming rigs. What the APU will allow is for better encryption/decryption (an example is Intel's integrated AES on i5), as well as other FP type work. Graphics will benefit, but not that much.

I really expect to continue to see discreet graphics cards for gamers for the foreseeable future. However, that does not mean gamers won't benefit. Physics can then be easily handled by the APU.

In short, at least for the next decade, the APU will certainly make games better, but it won't replace the dedicated and discreet GPU.

 

[cbn]

I think you expecting to much from Llano in the area of graphics. Initially I would expect it to work only a little better than an IGP. So while the cost of the CPU will go up, the cost of the chipset will go down. Because the GPU is on die with the CPU, graphics performance will increase but not incredibly.

We have to hope AMD can fix any bandwidth deficiencies so the full 3D potential of the Llano GPU can be used.

In fact, I am hoping the power circuitry on the mainboards is sufficient enough to overclock both the Llano CPU and Llano GPU with a tower cooler.

With plans of revised Crossfire Architecture (allowing discrete Radeon and Lano GPU to work together) maybe this will be possible? Wouldn't removing 75 watts from the equation (power through PCI-E slot when discrete card is removed) allow more room in the VRMs for overclocking Llano APU as a standalone?

I really expect to continue to see discreet graphics cards for gamers for the foreseeable future. However, that does not mean gamers won't benefit. Physics can then be easily handled by the APU.

In short, at least for the next decade, the APU will certainly make games better, but it won't replace the dedicated and discreet GPU.

If a graphics improvement can benefit gamers, it can also benefit a lot of other non-gaming programs. But how to move kids out of games and into more productive programs as they get older? Having *some* educational and real-life basis to the games/puzzles will probably help a great deal with easing the transition. When a child sees first hand the problem solving capabilities of a powerful desktop (with generous LCD screen) in a simulation it might give them inspiration to specialize in a certain field.

Like I said before I am not in the IT field, but certainly most of us can agree upon the idea that computing will get more visual in the future.

 

[grimpr]

Xbitlabs - AMD’s Sabine Platform Takes Nearly Final Shape.

AMD Sabine: Up to Four Cores, 20W – 59W TDP, DirectX 11 Core, USB 3.0, Vancouver GPU

http://www.xbitlabs.com/news/mobile...Sabine_Platform_Takes_Nearly_Final_Shape.html

 

[grimpr]

New information for ATIs next gen CrossfireX bag of tricks.

Multi-GPU

Moving the Multi-GPU component of the Catalyst driver outside of the 3D driver does more than allow anytime CrossfireX application profiles. The new Multi-GPU driver is modular and common across all APIs, such as DirectX or OpenGL.

This enables asymmetrical CrossfireX configurations, and more than that is critical for the forthcoming AMD CPUs with integrated APUs. Technically it would be possible to combine an ATI Radeon HD 5800 series with an ATI Radeon 9700! So what we see being unleashed here is a paradigm shift in how CrossfireX and Multi-GPU scaling is working. Combining this with the DirectX 11 multi-threaded submission execution model to multiple devices gives a picture of a broad, parallel pipe for execution.

This is obviously a broad-ranging and forward thinking change. With this new Multi-GPU driver architecture, a lot more configuration possibilities exist. Alternate Frame Rendering could be adjusted such that a 4:1 ratio could be used between an Enthusiast and Premium product, and need not be the only mode available - split frame rendering could be implemented with one GPU processing all the verticies, and the second for all pixels. Non-3D workload (i.e. DirectCompute, OpenCL) workloads could be scaled appropriately, and dynamically.

Rage3D - Will the MultiGPU driver allow for split frame rendering, where a more powerful primary GPU renders most of the scene and the secondary GPU does less? Or, for Tessellation/DirectCompute offload?

Terry Makedon - Potential is available, no support yet - it will be on the roadmap; dynamic load balancing to balance 1.8 for rendering, 0.2 for physics (for example).

Rage3D - Under the new MultiGPU driver component will OpenCL on GPU's be supported in a CrossfireX configuration, or will it need to be disabled to leverage all GPU's? Will the new MultiGPU driver component make any difference for existing GPGPU applications like F@h?

Terry Makedon - Scaling support doesn't yet exist but we hope to include in future.

This opens some exciting configuration choices, and allows users to continue to use their existing hardware with future generations - a nice return on their initial investment in AMD hardware. If future products allow shared GPU memory communication (for example, through the reintroduction of the Sideport), multi-GPU scaling and power efficiency gains could be substantial in an AMD platform equipped with next generation CPU + APU, and Radeon.

http://www.rage3d.com/articles/ati_catalyst_10/index.php?p=3

 

[cbn]

http://forums.anandtech.com/showpost.php?p=29551616&postcount=34

http://forums.anandtech.com/showpost.php?p=29551687&postcount=38

Here are two posts that came up during a recent discussion of Fermi's and Cypress memory bandwidth.

So if the PC gaming industry adopts a lower quality model being tessellated out to a higher quality model it looks like a GPU sharing system memory might work after all.

However, if the trend goes the opposite direction the situation would be different. High quality model is used as the starting point with tessellation being added in doesn't reduce memory bandwidth requirements.