Many-core processors

Without taking into account graphics processing units (GPUs), which constitute a class of architecture in their own right, large-core processors differ from multicore processors not only in the number of cores, but also in a number of other characteristics: core type and performance, hierarchical decomposition into clusters (or nodes) of cores, memory model (shared or distributed memory) and the software problems associated with the fact that they are almost always used as coprocessors. These manycores are used in two main classes of applications: those, such as scientific computing, where high performance is the most important criterion, and high-performance mobile and embedded applications where energy constraints are fundamental.

The various characteristics are presented, along with the different variants. Four examples of manycore processors are then detailed. The first two are designed for high-performance computing and dissipate two to three hundred watts:

• Intel Xeon Phi processors and coprocessors with Knights Corner and Knights Landing models;

• the manycore SW26010 used in the Chinese supercomputer TaihuLight, which from June 2016 to November 2017 topped the TOP500 list of supercomputers.

The following two examples are architectures designed for high-performance mobile and embedded applications and dissipate from a few watts to three tens of watts:

• processors implementing Adapteva's Epiphany architecture, with two versions in use at 16 or 64 cores and a 1,024-core version that appears to be a failure;

• Kalray's MPPA architecture, including a study of MMPA2 and MPPA3 versions.

The Xeon Phi approach, using shared memory and cores that are not sufficiently simplified compared with multicore cores, has insufficient performance compared with graphics processors, leading to its abandonment by Intel in 2018. The SW26010 and Pezy SC-2 manycores use core hierarchization and shared memory. They are used in some of the world's highest-performance supercomputers. Manycore processors for high-performance embedded applications are promising. While Adaptera uses only the 16- and 64-core Epiphany versions, Kalray, with its MPPA3 architecture, whose power dissipation remains around 20 W, and with its arithmetic extensions for computer vision and deep learning, seems well positioned in this class of applications.

Manycore processors are not the continuation of multicore processors with a greater number of cores. The number of cores is nowhere near exponential growth. New artificial intelligence applications seem to favor low-power manycores and...