NVIDIA enables GPU System Processor of some data center computing cards
The full name of GSP is GPU System Processor, which is a GPU system processor. As recently shared by Nvidia in the Linux documentation, this is a new feature hidden in Turing or newer architecture GPUs. Its role is to shift the driver’s responsibilities from the CPU to the GPU, such as GPU initialization and management tasks, to improve performance and efficiency.
Recently, Nvidia has officially unlocked it with the latest driver, and the GPUs involved include the Tesla T4 based on the Turing architecture to the A100 based on the Ampere architecture. GSP can be enabled by adding NVIDIA driver version 510.39.01 or later if the user is using an eligible compute card. Of course, users can also choose to forcibly disable GSP, but some display or power management-related functions will not work properly.
A module of the GSP on the GPU that acts as a coordinating processor for the GPU. The functionality of the GSP is still limited at this stage, but Nvidia says it will continue to expand in future drivers. Perhaps in heavy GPU-centric workloads, GSP can play a key role. There is speculation that the GSP is based on the RISC-V Falcon microcontroller which was introduced by NVIDIA back in 2016
Nvidia has now identified this feature in enterprise solutions, and it will also be available on Max-Q laptops launching in 2022 and beyond. A recent video released by Nvidia, explains how the latest version of Max-Q maximizes the efficiency between the CPU and GPU, including the GSP, but Nvidia calls it a “command processor.” TomsHardware believes that, according to Nvidia, the so-called “command processor” and GSP are the same thing and have the same way of working.
There is not much information about GSP at present, such as help or advantages for the actual performance of the GPU, but I believe that NVIDIA will continue to develop this function.
