Can AMD Ignite Fusion?

June 7, 2010

AMD is the perpetual also-ran in the CPU and GPU (graphics processing unit) sweepstakes. According to AMD, their next big thing is Fusion, in which a CPU and a GPU are combined on a single chip, or APU (Accelerated Processing Unit). Is it possible that in 2011 AMD will make serious profits from this innovation? Or is this more like the power from nuclear fusion program, which after decades of work and billions of dollars of investment has yet to make commercially available electricity?

In my April 17, 2010 story, Paint AMD Black?, I discussed AMD's latest financial results. The second quarter ends with June, and that readout will tell us how well demand is holding up. But the real value in AMD stock, or lack thereof, is in whether it has made the right choices in its roadmap, given its difficult position versus rivals Intel and NVIDIA. While there are other aspects to that roadmap, Fusion is probably the most important in the 2011 time frame.

There are two major forks in the computer processor road in the 2009-2011 time frame. One is about graphics processing, including using processors designed for graphics to compute other types of problems. The other fork is between traditional PC processors and mobile processors, where low power requirements are as important as computing and graphics capabilities.

Intel is the largest of the three rivals, so I will describe its known roadmap first, then contrast that with the roadmaps of AMD and NVIDIA. Intel dominated the market for PC computer chips almost from its inception, but it has traditionally supplied only minimal graphics capabilities. Keep in mind that a regular CPU can do graphics calculations. A GPU speeds up graphics calculations. GPUs can do this because graphics calculations use a relatively small number of well-known algorithms to accomplish their tasks, and these algorithms can run in parallel, all at the same time. In addition to letting the CPU do graphics, as that became unsatisfactory even on low-end computers, graphics were added to motherboard chip sets. This is called integrated graphics, and Intel (as well as NVIDIA and AMD) developed low end-graphics for motherboard chip sets.

Intel also tried to develop what are called discrete graphics chips. These are not on a motherboard, but on a separate card, called a video or graphics card. Intel recently abandoned its Larabee (see Anand's Thoughts on Intel Canceling Larrabee) project for a high-end discrete graphic chip, but will be putting its expertise to work by adding the failed Larrabee circuitry to its own CPUs, some time in the future. So right now Intel has no clear strategy for dealing with GPU computing on desktop or server computers. If you want that, you need to add an NVIDIA or AMD ATI GPU card to an Intel-based machine, and we are already seeing a lot of that.

In the mobile space Intel sold its floundering mobile application processor unit to Marvell, which in retrospect was a mistake, as Marvell is quickly becoming a big player in mobile devices. Instead Intel concentrated on Atom, which was originally used in the netbook space. Intel's strategy is to make Atom more powerful, yet less energy needy, to move more into the smartphone and tablet computer space.

AMD has decided to focus on the PC (desktop and notebook) and server market with graphics capabilities being a major weapon to win market share from Intel. The last time AMD won a lot of market share from Intel was also due to a roadmap decision, when Intel tried to bifurcate customers into 32 bit and 64 bit users. AMD instead integrated 32 bit and 64 bit capabilities on its Athlon and Opteron chips, which was so popular that Intel had to change it roadmap.

Will it work this time? One problem is that AMD is typically a half-step behind AMD in the process technology it uses to manufacture its chips, so it either needs a larger chip to accommodate the same number of logic gates, or must make due with fewer gates on a similar sized chip. GPU chips for high-end discrete graphics are big, and so are the CPUs for PCs. Combining both on a single chip could lead to a lot of defects, low productivity, and negative margins. The first Fusion chips will probably have the equivalent of mid-range GPUs mixed with mid-range CPUs. Yet if defect rates are low enough, these could be formidable chips. There would be no need to run to an external motherboard graphics chip or discrete graphics chip in order to do the processing. For many applications that lack of communications lag will make up for the lower processing power of each unit.

AMD has simply decided not to compete in the mobile space except to the extent that its notebook processors are able to pick up some of the netbook or tablet market. Given the importance of the mobile market, that may not seem smart. But the mobile market is fiercely competitive; even Intel has floundered there, against the likes of Qualcomm, Broadcom, Marvell, and many others. AMD has limited resources compared to Intel. It can't afford to enter markets where it has no competitive edge.

NVIDIA has an interesting strategy too, and it involves a different take on each of the forks. It is going heavily into mobile with its Tegra processor. It is ahead of AMD in high-end graphics, and graphical computing, with its Fermi and CUDA technologies. Clearly NVIDIA would like to allow their GPUs to also do the work of CPUs and thereby crash the PC motherboard party. They could do that by adding ARM based processors to Fermi. But it would be a whole new marketing world for them. They might simply wait and watch the ARM/mobile space eat up the old PC space the way PCs once cut into the minicomputer space.

I own AMD stock and think it is currently undervalued, but there is no denying that Intel is much larger that AMD and has lots of profitability, while AMD has just squeaked by for years. Until AMD shows several successive quarters of profits and market share gains, and it is clear that it chose the best roadmap, I don't expect much short-term upward movement of the stock. On the other hand, waiting until all the good news is in means missing out on almost all of the stock appreciation, if it happens.