Radeon Linux Driver Adds Option To Limit Number Of Enabled CUs
Longtime AMD open-source Mesa developer Marek Olšák introduced this new AMD_CU_MASK= environment variable override as a way of controlling the number of CUs enabled -- to artificially limit the number of compute units enabled -- for both RadeonSI and RADV drivers.
AMD_CU_MASK can be set to an ID list of the compute units to enabled or a list of the compute units to enable, with syntax format examples such as "0x337F" or "0,2-4,7".
This AMD_CU_MASK override was requested internally within AMD while may have some usefulness externally as well for those wanting to analyze the performance impact of lower CU counts and analyzing how different workloads behave or scale with varying compute unit counts, etc.
See this merge request, which was merged into Mesa 22.0-devel today, if interested in this ability for limiting the number of enabled CUs for Radeon GPUs.