Calculation, memorization and communication

In this section, we focus on the computational aspect of neural processors. The cores of these processors are intended to provide the computational power needed to execute convolutions and fully-connected levels. But computational performance implies that the various neural cores are interconnected via high-performance networks, and that memory subsystems are capable of feeding them with data.

• Neural processors are multicores interconnected on-chip or on-chassis via high-performance interconnection networks, generally of the 2D grid type.

• The memory bandwidths required are considerable. To take just one example, Intel's 2020 Xeon processors (E7-8890 v4, E5-2699A v4 with 24 and 22 cores respectively) have a maximum memory bandwidth of around 80 GB/s. By comparison, the memory bandwidth of the Xilinx VersaI Core neural processor...