Concluding remarks

Hardware developments for deep neural networks are constantly evolving. They cover a wide spectrum, from data centers to systems for the Internet of Things and mobile devices (smartphones, tablets). We have only considered hardware from major operators (Google, Microsoft, etc.) or chip suppliers (Arm, Intel, NVidia, Xilinx, etc.). There are a large number of small companies and startups offering neural network circuitry.

While it is still difficult to characterize them precisely, a number of trends can be identified:

• neural processors, particularly at the top end of the market, feature a large number of cores interconnected by high-performance networks and fed by high-performance memory systems. They are closer to very large-core processors (presented in