Inside Intel: From Silicon to the Worldby Anand Lal Shimpi on February 11, 2002 3:58 AM EST
- Posted in
Multi-core CPUs: The
Over the past decade the focus in improving CPU performance has been on extracting as much instruction level parallelism (ILP) out of code as possible. The idea behind ILP is to execute, in parallel, as many instructions as possible. This keeps execution units busy and it makes computing efficient.
There is only so much ILP that can be exploited before you begin to receive diminishing returns. This is why CPU manufacturers are looking towards ways of improving performance through thread level parallelism (TLP), or the execution of as many threads in parallel as possible. Both AMD and Intel are working on methods of improving TLP; AMD is rumored to be testing a multi-core version of their Sledge Hammer processor and Intel has more recently introduced Hyper-Threading Technology which allows for the simultaneous execution of threads on a single processor.
Intel’s labs are also hard at work at discovering the feasibility of a multi-core CPU. The “simple” approach would be to have two identical cores running in parallel in a single package, however as we’ve outlined in the past there should be much emphasis on the packaging part of the equation. Intel has been toying with an interesting, more elegant approach.
With an increased focus on power consumption, Intel’s engineers have been hard at work in finding ways to make sure that future generations of processors won’t require water or other extreme cooling measures. And having two cores running in parallel will definitely make power consumption and dissipation very big issues, very quickly.
Intel’s theory revolves around having two cores, but not necessarily of identical characteristics. It turns out that a good part of the time your CPU is waiting for things like your disk drive or main memory and thus isn’t gaining much from running at full speed. Thus if you had two cores, one high-performance and one lower-performance you could save power and potentially even transistors if some functionality was removed from the lower-performance core. The high-speed core would be used for critical data execution and the other for execution of data that more dependent on slower paths such as from main memory or disk subsystems. Obviously there have been no products announced to feature this type of technology but Intel’s early findings show that the technology is promising. Remember that the labs aren’t product driven, rather technology driven so they’re not working on any particular products per se, just ideas that may end up taking shape in future products; it’s interesting nonetheless.