In my previous post, I linked an article about an interesting approach to reuse the heat produced by a supercomputer for saving on the bill of a university.
In this article in IEEE Spectrum, instead, the authors present an approach to building supercomputers that consume less. The idea is very simple but very effective: instead of using a lot of powerful multicore processors in parallel, the authors plan to use much more processors, but each one less powerful.
In the past, supercomputers were built by designing everything from scratch for maximum performance. However, this approach is today too costly, and it does not scale very well. Today, many supercomputers are built by assembling together existing commercial processors. Hower, processors used in PCs today consume a considerable amount of electric power.
Now, the crucial observation is that processors used in smart phones have a considerable computational power and consume much less energy (because they have to work on batteries). In other words, they are more energy efficient. Thus, why not building a supercomputer with the processors used in smart phones?
Indeed, it seems a very good idea. But, how much do we gain?
The authors of the article make an example of an application for weather climate calculations. To improve the current models on weather, researchers need much more powerful supercomputers than available today. Such supercomputers can be built using a commercial PC-class processor, as the AMD Opteron; or with a smart phone class processor, like the Tensilica XTensa LX2. In the following table (taken by the article) there is the comparison between the two approaches:
|Processor||Clock speed||Gflops/Core||Cores needed||Power|
|AMD Opteron||2.8 GhZ||5.6||1,700,000||179 MW|
|Tensilica Xtensa LX2||500 Mhz||1||10,000,000||3 MW|
As you can see, the total power used by the embedded processor supercomputer is more than 50 times inferior to the power used by the supercomputer built with the Opteron!
Of course, a supercomputer is much more than 10 million processors. The interconnection and the memory is also very important! However, recently a lot of research on GRID platforms has beed developed both in Europe and in the US. I think that building such a supercomputer as connection of 10 millions small systems running Linux plus some GRID platform should not be too difficult. I can’t wait to see it working!