AMD, Intel Race To Merge CPU and GPU

The graphics processor continues to evolve, not only in terms of hardware design, but also in its usage model. Both Nvidia and AMD have pushed for a larger application base for their GPUs, thereby creating a wider range of potential uses (and customers) for what's known as a GPGPU (general-purpose graphics processing unit).

Tools like OpenCL, DirectCompute, and Nvidia's proprietary CUDA all allow for computing tasks on a dedicated GPU, giving programmers access to the power of the GPU to perform non-graphical duties such as video transcoding. The logic behind this is sound, as many high-end graphics processors go mostly unused when performing standard office duties.

Ball of Con-Fusion?

But if step one is to offload some CPU tasks to the GPU, step two is to erase the distinction between the two processor types altogether. This is what AMD has been touting for some years under the name Fusion: combining the features of a CPU, GPU and Northbridge into a single die, which the company has dubbed an APU (Accelerated Processing Unit) and promises to ship late this year or in early 2011.

According to AMD, an APU is not just a CPU grafted onto a GPU, but a next-generation processor combining programmable CPU and GPU cores, along with memory, I/O and video controllers. Future revisions will blur the lines between processor types even further, with shared resources and multi-use cores.

While it's fun to coin new terms, the Fusion's x86 processor will remain mostly unchanged from current offerings, with a dual- or quad-core Phenom architecture on a 32- to 40-nanometer process. This is a bit of a disappointment, as Fusion was expected to herald a new CPU architecture, but we will have to be satisfied with a die shrink.

The architecture of the Fusion APU includes the x86 cores and a SIMD (single instruction, multiple data) Engine Array -- traditionally called the GPU -- along with an integrated memory controller and bus joining the two together. The single memory controller for both CPU and GPU access could be problematic, and even AMD admits this challenge in its documentation.

The APU idea also changes other computing terminology. Instead of CPU and GPU tasks, AMD now refers to scalar and vector processing, respectively. The traditional scalar operations are taken care of by an x86 CPU, while the graphics processor handles any vector operations (few as they may be for office apps). While it's mostly a play on verbiage, the merging of both types of calculations does highlight another facet of the Fusion brand name.

The Fusion will ship in two flavors, "Llano" for desktops and "Ontario" for notebook platforms. The Llano platform will sport dual- and quad-core models with some form of ATI Radeon HD 5000 graphics, while Ontario will be a dual-core processor with an extremely low TDP, reportedly able to scale down as low as 1 watt.

Intel's Preemptive Strike

Of course, while AMD has been talking up Fusion, Intel has not been idle. In fact, the silicon giant has beaten AMD to market with an integrated CPU/GPU, having released Core i3 and i5 "Clarkdale" desktop and "Arrandale" mobile processors with integrated graphics early in 2010.

Historically, Intel has taken part in several product launches with accelerated release schedules designed to head off any AMD competition. Perhaps the most (in)famous, the Pentium D, was Intel's entry in the race to ship a dual-core CPU -- two Pentium 4s thrown together in a multichip design that did not include all the hardware on one die, something AMD accomplished with its subsequent Athlon 64 X2. The Pentium D may have been first, but the Athlon 64 X2 was the better processor.

Today's Clarkdale architecture includes a multi-core CPU, integrated memory controller and onboard GPU, but in a multi-chip format using 32nm CPU and 45nm GPU dies joined by an external bus. This is exactly what Intel did with the Pentium D, putting two processors on the same package, and it comes with the same caveats, namely lower memory bandwidth, duplicated resources and less efficient performance.

On the other hand, AMD has stated its upcoming Fusion release will be a fully integrated chip with on-die CPU and GPU functionality. This difference may seem slight, but can affect everything from performance to production efficiencies. By incorporating a single integrated die with all of the functionality, AMD's Fusion could be faster, smaller, more energy-efficient, and cheaper to produce.

If history is any indicator, AMD may be poised for a comeback. The fully integrated, one-chip Athlon 64 X2 emerged to great fanfare, literally burying the Pentium D. In fact, that was the last golden age for AMD, when the company could thumb its nose at Intel in terms of performance, value and even production technology.

Making it in the Mobile Market

While entry-level desktops can certainly benefit from the cost savings of an integrated CPU/GPU chip, the largest market will likely be in mobile devices. Intel's Arrandale Core i3/i5 processors have already brought this to laptops, but as noted above, this is hardly the best implementation of integration.

Graphics performance is becoming more important in the notebook market. Desktop sales are down, while laptops are up, and there is no question that both consumers and businesses are replacing traditional PC towers with notebooks.

In terms of desktop replacement capability, standard Intel integrated graphics simply won't do, and most power users go with a dedicated GPU. If AMD can ramp up the graphics performance of its Fusion APU, it could ride the wave of the mobile explosion.

Another potential angle has AMD looking at the rapidly emerging tablet market to push its Fusion technology. As a single-die, monolithic design, the Fusion APU will have performance, energy and heat advantages over current Intel designs, and can therefore be used in smaller form factors that still require CPU-plus-GPU power.

The iPad is a neat little toy, but if the market grows beyond this into full-featured slates that can actually replace a notebook, then it will require a more robust processor that excels in power savings. If this market takes off, AMD could go along for the ride.

Can Fusion Beat Intel?

With AMD poised to drop the Fusion bomb, and Intel shipping an old-school, multi-chip counterpart, the timing seems right for AMD. But Intel has learned its lesson and now concentrates on being first to market, then following that up fast with the real deal -- in this case, the upcoming "Sandy Bridge" platform, which will feature a 32-nanometer-process fully integrated CPU-plus-GPU chip.

If the Clarkdale and Arrandale bundles can hold Intel over until Sandy Bridge emerges, then AMD is in a lot of trouble, especially if the release dates are even remotely close.

Matching Intel's product release schedule is the major obstacle for AMD, as it seems that whenever a new AMD product is on the horizon, Intel can just reach into its bag of tricks and hit the market first. The product may not be an elegant implementation, but it does serve as a stopgap method for Intel to buy some more time.

The X factor in all this is the amount of GPU power AMD will give the Fusion. Intel's Clarkdale GPU is like a dancing bear -- it's remarkable that it can perform at all, but it's not going to win any competitions. AMD needs to set its integrated GPU goals a bit higher, and rumors are swirling that it could be anywhere from an entry-level Radeon HD 5450 equivalent all the way up to a more mainstream HD 5670, or even HD 5750, solution.

Choosing the right GPU component for Fusion is a difficult balancing act, as AMD still relies on sales of ATI graphics cards and there is a risk of cannibalizing them if the Fusion graphics are a bit too powerful. So no one is expecting anything high-end. On the other hand, Intel has little to fear from an integrated Radeon HD 5450. Going too low on the graphics scale could make Fusion a fizzle.