AMD's Fusion: Revolution or dangerous adventure?

Sunnyvale (CA) - AMD today confirmed that it will be developing a processor, code-named "Fusion," that will combine a traditional CPU with graphics capabilities under one roof. If the deal was really just about integrating graphics, ATI could turn out to be a tremendously expensive deal. But in the end, Fusion may not be about graphics at all - but about a technology that could revolutionize the CPU.

A simple look at AMD's balance sheet reveals that ATI isn't just an acquisition that will be adding some features to AMD's existing line of products. The purchase price of $5.4 billion represents more than one year of AMD's revenues and more than three years of the current profit run rate. When AMD got in the final states of merger talks, the company stock was in a much better shape than it was today, which made ATI somewhat affordable, given AMD's cash reserves, which stood at about $2.3 billion on 1 October of this year.

To finance ATI, AMD was a forced to take out a $2.5 billion loan, use more common stock (58 million shares for a total value of just under $1.2 billion) - which means that the company had to shell out about $1.7 billion in cash - more than 70% of its cash reserves. This may not have been a simple decision with more than 15,000 employees on board and combined profits of less than $200 million per quarter in an industry that could dramatically shift within a few months. And some may consider this scenario as too risky, if AMD simply wanted to add a few graphics cards and chipsets to go along with its processors.

The announcement of Fusion provides a first idea what AMD may have in mind with ATI: The central value of ATI may not revolve around graphics at all, but about integration talent as well as a concept that is currently evolving under the term of "stream computing" - a technique that make use of the enormous floating point capabilities in applications other than graphics.

Hal Speed, marketing architect for AMD, explained in a conversation with TG Daily that the concept of Fusion is similar to "Hammer," which combined the memory controller with the CPU in AMD's K8 processors. "Think of it as an umbrella, a concept or vision for integrating the CPU and the GPU into one piece of silicon," he said. "Fusion is the umbrella initiative with a potential that is not limited PC, servers and mobile systems, but extends to anything that has a CPU and GPU working together." According to Speed, Fusion will be targeted at mainstream and low-end computing - as long as graphics are concerned - initially. "Gaming enthusiasts will be better off with a discrete graphics solution or even multiple graphics cards. Of course, we will continue to make Radeon graphics cards and service those markets," Speed said.

In that view, Fusion will allow AMD to pursue a higher level of integration than we see today: The graphics capability of the chipset joins the already integrated Northbridge in AMD's processor and leave the Southbridge - as only remaining external element of the chipset - on the motherboard to handle I/O tasks. This strategy could allow AMD, for example, to offer potentially smaller form factor components that are easier to integrate for system builders such as Dell.

However, graphics chips aren't just used for graphics anymore. A while ago, ATI discussed the concept of load balancing, which lets the company shift GPU processing power especially in Multi-GPU Crossfire configurations to handle other applications, such as physics processing. In January of this year, ATI said that a third graphics card could be used in enthusiast systems to cover that task. More recently, Stanford University's Folding@Home project announced that it will enable the use of graphics cards within its distributed computing project - for an expected 500x increase of computing horsepower. ATI's X1900 and X1950 graphics cards currently unleash a floating point performance of about 375 GFlops - more than ten times what dual-core desktop and workstation processors can offer today.

So, can this mainstream CPU potentially be fine-tuned to become a high-end processor that opens a whole new world of possibilities for software developers? Speed answered our question with a brief "yes." But the problem with this idea really is that floating point capability isn't what is required by the mainstream applications we use today. Speed conceded that the trend is "more in the technical community and academia, where researchers code algorithms that take advantage of the vector processing capability in GPUs." AMD hopes that software will evolve over time to tap into the unused potential of a CPU that houses a powerful GPU. Down the road, Fusion could accelerate security and cryptography algorithms and more mainstream applications such as image and video processing. If AMD has its way, we could be standing just before a whole new world of computers - systems that reset our perception of what computers and applications can do. And even before such applications become available, Fusion could also end up in a high-powered processor for scientific cluster installations that will multiply system performances before the end of the decade.

Speed declined to discuss details just how far Fusion could go and told us that AMD will shed more light on its plans during its analyst conference, which is scheduled to be held on 14 December. However, we learned that Fusion may not replace AMD's standalone processors: "I do not think that GPU functionality will become a standard feature in our CPUs."It is probably too early to tell and we will see how software will evolve. Based on the horizon I am seeing, there is still a need for discrete CPUs and GPUs."

In any case, ATI has a long way to go and is facing one of its biggest and most important challenges yet: It has placed a big bet on a next-generation processor that does not only need to close the gap to Intel, but aims to provide - on numerous fronts - a lead over its rival in Santa Clara. We would not only be surprised to see a Centrino-like mobile platform in the 2007 time frame, but an integrated CPU/GPU technology that can be marketed as entry-level computing solution on the very low-end of the marketed as well as a high-end computing solution in servers that cost multiple millions of dollars.

The question will be how AMD's existing and potentially new customers react to this idea and if the software community will respond to processors that do not only need new algorithms but can also handle many more threads than today. And, of course, there is Intel - which may or may not have to react to AMD's ideas.