The Core of the Problem

Chances are that if you have purchased a computer in the last three years, it has more than one core processing unit.  About 5 years ago, processor designers hit a brick wall.  They could not make CPUs run any faster.  Yet, ever-reducing transistors sizes kept providing more transistors in the same chip area.  The decision was to begin placing more than one processor on a chip, or multicore.

Intel's Barrett leaves chipmaking legacy, shortfall reviews some of the history.  Originally the Pentium 4 family was envisioned to operate up to 10 GHz.  But faster clocks result in much increased power consumption and heat dissipation.  The Intel Netburst architecture just ran too hot.  To save the company, Barrett's brain trust decided to place two CPUs on the chip and divide the workload in half.  Sounds reasonable, doesn't it?

It only works if the problem can be partioned, or vectorized, as described in older computer architectures.  Even if it can be partioned, taking the problem apart and sending it to multiple CPUs and then bringing all the results back together at the exact right moment in time is a difficult problem.  When it comes to CPU cores, more not always better tells that on some key algorithms, Sandia Labs observed improvement with four cores instead of two.  Increasing to eight cores barely improved performance and sixteen cores actually performed like two cores!  The problem is the sharing of data between the processors, called memory bandwidth.

Hardware and even memory bandwidth changes, so the breakpoint today might be 4 processors, but tomorrow it will be much higher.  The really hard problem vexing computer scientists is how to produce software that can be partitioned to multiple processors.  It seems that data flow processing is like nuclear fusion - the breakthrough will be next year.  It is a really, really, hard problem - trying to determine how to break the calculations into pieces and then schedule how everything comes back together again.  Multicore processors outpacing key business software tells that today's top-end servers have eight processor chips with eight cores, resulting in 64.  Yet today's business software cannot scale to keep that many processors busy.

Dual or quad core - which is best for you? answers a question that many new computer buyers have.  Should I buy a quad core or a dual core?  The columnist recommends a dual core for most enthusiasts.  Something else surprising - Tests show Win XP still fastest for multicore tells that the venerable Windows XP, developed before multicore processors is better at partitioning loads than Vista or even the new Windows 7.  The reason - XP is smaller and lighter than the newer operating systems.