multicore programming, and what will not work

We wrote about XMOS Semiconductor about a year ago when the company first emerged from stealth mode and introduced its programmable semiconductor technology called Software-Defined Silicon (SDS). Essentially, the idea was to configure the underlying hardware to handle the software tasks at hand via a high-level description language. Obviously, there are many ways of solving parallel programming problems for multicore processor, however, according to Professor David May who is the CTO of XMOS, many approaches to date will simply not work, as he recently revealed in a column by David Manners over at ElectronicsWeekly.com.

Professor May is mostly critical of Intel and Microsoft, arguing that the basic definition of their problem is flawed in that they are trying to take current sequential applications and enable them to run better on multicore processors – which is "virtually impossible." He further points out that the shared memory approach is flawed as well since every time more cores are added, they will end up competing to gain access to the same resource, essentially making an already complex problem even more difficult. He is also not a big fan of people who put their faith into compilers or abstraction layers for legacy software, arguing that the former take too long to develop and optimize while the latter are simply inefficient. One of the more interesting quotes from the article is with regards to computer engineer in general and a paradigm shift that David thinks is necessary: "A universal computer is an infinite array of finite processors, not a finite array of infinite processors." - Now you have something to contemplate over you next coffee.