The long-awaited patches for permitting the open-source NVIDIA “Nouveau” upstream Linux kernel driver to leverage NVIDIA’s GPU System Processor “GSP” firmware for dealing with GPU re-clocking and different {hardware} duties with RTX 20 GPUs and newer have been posted. With this set of 44 patches additionally comes the preliminary GPU {hardware} accelerated assist for the GeForce RTX 40 “Ada Lovelace” GPUs that’s constructed upon this new GSP driver code path.
Ben Skeggs of Purple Hat has posted the set of 44 patches offering the preliminary groundwork for enabling GSP firmware use by the Nouveau kernel driver. Whereas the NVIDIA GSP is current going again to GeForce RTX 20 sequence, for now the patches don’t use the GSP path by default aside from the newest RTX 40 sequence the place that’s the solely possibility.
With the GSP firmware chargeable for extra energy administration duties and associated {hardware} initialization, it is moved out of the Nouveau driver and in flip GPU re-clocking ought to be working. This in flip ought to present some efficiency advantages over the present Nouveau (horrible) efficiency on RTX 20 and RTX 30 sequence {hardware} however nonetheless different driver bottlenecks stay earlier than Nouveau might doubtlessly compete with the NVIDIA proprietary graphics driver stack.
This Nouveau GSP work can be presently tied to the R535 present secure sequence firmware. The NVIDIA GSP firmware is not essentially secure throughout driver variations, so we’ll see the way it performs out in observe.
With the 44 patches there’s then RTX 40 sequence accelerated assist when having the NVIDIA firmware information current in your system. For these on RTX 20 Turing and RTX 30 Ampere GPUs, the new.config=NvGspRm=1 kernel module possibility is what’s wanted to get the GSP driver path working.
Enabling the NVIDIA GSP-RM firmware use by the Nouveau kernel driver is round fourteen thousand strains of recent code. See the 44 patches if . I will be working some benchmarks as soon as it appears to be like like issues are pretty settled. With some luck maybe we’ll see this upstreamed for the Linux v6.7 kernel cycle.
