| Thursday, May 15, 2008 Login | Register | |||||||||
| ||||||||||
| Cell Processor: Faster or Hype | |||||
| Question | |||||
| Is the 'Cell' really all that faster than normal processors or is it all just hype? | |||||
| Answer | |||||
| In general, the Cell processor is all hype. In most cases, it's slower than a normal CPU. Cell is a general purpose processor, but for general purpose code, Cell is not ideal. For application specific processing, Cell is inefficient. It fails to deliver in both regards.
However, there are some cases where the Cell can shine. If a task requires a general purpose processor, can be easily split into many parallel independent subtasks, and has predictable memory usage patterns, then the Cell can perform better than other types of processors. Image analysis is one example. In other words, the Cell is a general purpose processor designed for specific types of code. Other similar many-core processors built primarily for parallel stream computing can be faster than Cell even in these specific areas. I'll explain also with terms unrelated to computers. The objective is to build coffee tables. There are three options. You can have the coffee tables built by: a very intelligent person with excellent skills in many areas, four dumb people who are well-trained at building coffee tables but not much else, or eight robots that each do nothing else but build one specific type of coffee table. If you just want a very limited number of types (1-8) of coffee tables produced as fast as possible, the robots are the best solution. If you want many types of coffee tables (9+), the four dumb people are better suited. If you also want to accomplish many other things besides just building coffee tables, the intelligent person with many skills can handle it all smoothly, while the four dumb people look at each other confused and flip through a How to book for Dummies. The one intelligent person is a normal CPU. The eight robots are dedicated application specific processors. The four dumb people are the Cell.  Date: 2008-04-23 | |||||
|