Taking the latest Nvidia Ampe microarchitecture as an example, the computing part of the GPU microarchitecture consists of a Stream Processor (SP),
a Texture mapping unit (TMU), a Tensor Core, a Ray Tracing Unit (RT Core), and a Rasterization Processing Unit (ROPs).
GPU's API (Application Programming Interface) plays a role as a bridge between applications and video card drivers.
GPU video memory is used to store the rendering data that has been processed or will be extracted by the video card chip,
and it is one of the indispensable core components for the normal operation of GPU. GPU video memory can be divided into independent video memory and integrated video memory.
At present, GDDR3, GDDR5, GDDR5X and GDDR6 are mainly used for independent video storage, while DDR3 and DDR4 are mainly used for integrated video storage.
The server GPU prefers to use HBM video memory in the form of chiplets to maximize throughput. The integrated video memory is limited by the 64 bit operating system.
Even if it is composed of two or even four channels, there is still a considerable gap between the bandwidth of the integrated video memory and that of the independent video memory.
This usually results in the performance of independent GPUs being better than that of integrated GPUs.