As reported in a recent Nvidia document for Linux, the company has quietly unlocked a new feature in its enterprise and consumer GPUs that has been hidden since the Turing generation. Known as the system processor GSP or GPU, this chip offloads driver functions from the CPU to the GPU to improve performance and efficiency. It has been officially unlocked for use with the latest Nvidia drivers.
This feature is available now if you have a suitable Nvidia enterprise GPU and Nvidia driver 510.39.01 or later. GPU support ranges from the Turing-based Tesla T4 to Nvidia’s A-series such as the A100, A2, and A40.
The GSP is a real piece of silicon on the GPU (like the NVENC controller) that acts as a co-processor for the paired Nvidia GPU. Its function is to offload GPU initialization and management tasks from the CPU directly to the GPU to reduce overall system latency and improve system performance. Nvidia says that GSP functionality is limited at the moment, but will continue to expand in future driver releases. So expect Nvidia to rely more and more on this co-processor to keep the CPU load down.
We could potentially see some performance improvement in both games and graphics-intensive applications where the CPU plays a critical role in feeding the GPU information. This is especially true now that modern games are taking advantage of the CPU more than ever before thanks to the thoughtfulness of lower-level APIs.
Even though Nvidia has stated that it has blocked this feature for enterprise solutions (for now), Nvidia has also announced this feature as a critical component for laptops. Max Q, which will appear in 2022 and beyond.
In a new YouTube video released today, Nvidia explains how its latest version of Max-Q delivers the best efficiency between CPU and GPU. One of these features includes the GSP, but in this case Nvidia calls it the “command processor” instead of the system’s GPU.
We believe that GSP and “command processor” are the same as they function in the same way, offloading low-level tasks from the CPU to the GPU in order to improve performance. In the case of Max-Q, Nvidia gives an example of an instruction check that performs a pointer check and a balance check, and how that workload would be shifted from the CPU to the GPU to improve performance.
We still don’t know much about the GSP, such as its real performance benefits or full functionality. However, we expect to know more soon as Nvidia continues to develop this feature in 2022.