Nvida generated some pre-CES excitement by announcing its Tegra K1 processor, which looks to be the most powerful graphics engine ever designed for mobile use. But the company also spread a lot of confusion by describing it as a “192-core processor” chip. How did we go from two- and four-core system-on-chip processors to 192 in one enormous leap.
Of course, we didn’t. There are cores, and there are cores. Specifically there are general purpose (CPU) cores and graphics (GPU) cores, and Nvidia was not terribly clear about the distinction in the announcement. The Tegra K1 in fact comes with either two 64-bit or four 32-bit CPU cores plus a 192 GPU core unit based on Nvidia’s much-admired Kepler architecture.
There’s a big difference between how these two types of cores work. CPU cores are designed to handle the general run of processing, with sophisticated units using technologies such as out-of-order processing and predictive branching to speed execution. Multiple cores work independently of each other and some CPUs can handle more than one process thread per core.
GPU cores use a very different architecture called simultaneous instruction, multiple data (SIMD). All of the cores execute the same instructions on parallel streams of data. This approach was, of course, developed for processing graphics where, for example, rotating an image requires performing the same mathematical operation on every pixel, a job that naturally lends itself to massive parallelism.
But GPU processing in recent years has moved well beyond graphics as it turns out there are many other computing chores then lend themselves to SIMD organization. A whole branch of computer science has arisen to take advantage of general-purpose computing on GPUs (GPGPU). Nvidia offers a set of programming language extensions called CUDA to help developers create GPGPU programs, while OpenCL is an open-source equivalent.
SIMD processing is not practical for every problem, but where it is good, it is very, very good. Nvidia should have been more precise in its nomenclature, but bringing a Kepler-class GPU to a mobile system-on-chip could create a new world of high-powered mobile computing.
