New Fusion Processor Architecture from AMD.

Seriously, now I totally understand why Intel is considered the world's largest fab, while AMD is considered a microprocessor company. It is simply due to focus. Intel has a large pool of manufacturing expertise, and it leverages this expertise in getting better products out the door. However, AMD has to leverage on it's design expertise to get products out the door.

Yesterday, AMD has announced that they are coming up with a new processor architecture, called Fusion [http://techreport.com/onearticle.x/11438]. The reason is simply because they do not believe that stacking on more cores will increase computational performance. I totally agree [http://sybreon.blogspot.com/2006/08/multicore-plateau.html]. Anyone with an understanding of computer architecture will know that, simply stacking more cores, is not scalable [http://sybreon.blogspot.com/2006/11/more-supervision-questions.html].

AMD has announced that it will be designing a processor architecture that uses what it calls XPU (Accelerated Processing Unit). Incidentally, I'm working on something similar for my project. It's just that I call it an ASPU (Application Specific Processing Unit). I'll have to give the AMD marketing department props for coming up with that TLA. Maybe I should adopt their term instead. I should keep abreast with this developing platform. Maybe it's possible to tack on my project onto it. Now, isn't that exciting!

So, what is the XPU all about? It's about off-loading specific computational operations onto dedicated "accelerators". This trend has been going on for a long time now. It started with graphics and sound, but has now gone on to encompass a lot of other computationally intensive tasks (physics, artificial intelligence, and what nots). So, the general concensus is that our venerable general purpose processor is going the way of the dodo.

In order to understand the reason why, we should think about it in economic terms. In the old days, cost of silicon real-estate was high and design tools weren't that great. Therefore, it made sense to design a machine, that is suited to do all kinds of tasks (ie general purpose). It could therefore be deployed it various applications/markets. However, it is now quite cheap to design stuff in silicon due to improved design/manufacturing processes. So, it now makes economic sense, to do quite the opposite.

An additional reason is because of software trends. Our applications today are much more computationally intensive than they were before. A general purpose core is typically larger (and hence costs more) than a specialised one, while delivering similar or lesser performance. So, it starts to make more sense, to build a chip with multiple different specialised cores, dedicated to performing specific tasks extremely well. I'd detected this trend a while ago. Hence, why I'm working on just such a topic for my research project.

Hmm, I wonder if AMD would be willing to give me a job with them after I finish. It would be interesting to design a future processor, without a single general purpose core in it.