Computer architecture : CPU and coprocessors/accelerators

Coprocessors have been around since the early days of computer manufacturing, even before the birth of integrated circuits. Their function was to accelerate program execution. In recent years, often renamed gas pedals, they have taken on new relevance for artificial intelligence (AI), particularly for PCs and data center servers. Today's circuits use several different types of gas pedal alongside high-performance multi-core CPUs.

The article Evolution of computer architecture [H 1 058] , published in 2016, examines the articulation between semiconductor technology, architectural concepts and the needs of major application classes. It focuses on the evolution of processors (CPUs). This article is still relevant today, even though AI applications have become increasingly important.

In this article, we summarize the general characteristics of CPUs, and present the features of the main types of gas pedal, and why they appeared and were, or are, used.

We present and discuss the following types of gas pedals:

• math and I/O coprocessors, used in mainframes and early microprocessors.

• coprocessors in the form of reduced CPUs specialized for one type of application, such as neural processors (NPUs) in AI PCs;

• coprocessors using a calculation model different from that of CPUs, such as GPUs and FPGAs;

• coprocessors using special physical phenomena, such as quantum processors (QPUs).

For a coprocessor/accelerator to be widely used, a number of conditions must be met:

• it must provide a significant performance gain over the CPU, of one or more orders of magnitude;

• the hardware interface between CPU and gas pedal should be standardized as much as possible;

• the software interface (SDK) must be mature and easy to use, such as CUDA for NVidia GPUs.