It’s been about 10 years since multicore processors burst on the scene, and we’re now seeing several innovative variations. At first, chip designers simply replicated CPU cores, filling their silicon with copies of the same brain. Now they are exploring alternatives—and these variations will change the way we benchmark performance and compare processors.
A basic symmetric or homogeneous design has two or more CPU cores of the same type running at the same clock speed. As the number of cores grows, however, this approach becomes less power-efficient, and the software struggles to use all the cores. So AMD and Intel introduced “turbo” technology that can run the CPUs at different clock speeds, which saves power and scales performance with demand.
AMD’s Bulldozer processors have pairs of cores that share some internal resources. They are similar to Intel’s multithreaded cores, and both techniques will continue to evolve. Some embedded processors have four or more threads per core.
To conserve power in mobile phones and tablets, ARM and Marvell have designed asymmetric or heterogeneous CPU clusters. ARM calls its strategy “big.LITTLE”—a big core sleeps while a little core handles routine chores. When the user needs more performance, the big core awakens and helps little brother. Someday, PC processors will do this, too.
Freescale and Texas Instruments have chips that combine conventional CPU cores with a digital-signal controller (DSC) or digital-signal processor (DSP). The CPU runs application software while the DSC or DSP controls a motor or decodes wireless signals. PC processors could do this, too, although DSP instruction-set extensions may be an acceptable substitute.
As these and other variations proliferate, processor benchmarking and comparisons will become more difficult. But for PCs, smartphones, tablets, and other devices running highly variable workloads, the power efficiency of asymmetric multiprocessing will override other considerations.