In a turn of surprising events, an NVIDIA engineer pushed a Linux kernel correction, resolving a regression of the performance observed on integrated and dedicated AMD equipment (via Phoronix). It turns out that the same engineer inadvertently introduced the problem in the first place with a set of modifications made to the nucleus last week, trying to increase the space of the PCI bar to more than 10tib. This ended up wrongly reporting the GPU as limited and empowering performance, but fortunately, it was quickly picked up and fixed.
In the open-source paradigm, it is an unwritten rule to repair what you break. THE Linux core is open-source and accepts everyone’s contributions, which are then examined. Responsible contributors should help solve the problems resulting from their modifications. Thus, despite their rivalry on the GPU market, FOSS (Software Open Source) is an avenue that folds the abyss between AMD and NVIDIA.
The regression was caused by a commitment which aimed to increase the space of the PCI bar beyond 10tib, probably for systems with large spaces of memory. This has indirectly reduced a factor called kaslr entropy on consumer x86 devices, which determines the random place where the core data is loaded in memory on each start for security fine. At the same time, this also artificially inflated the beach of the accessible memory of the nucleus (direct_map_physmem_end), generally at 64tib.
In Linux, memory is divided into different areas, one of which is the area of zone which can be associated with a GPU. The problem here is that when the kernel initializes the zone device memory to GPUS RADEONAn associated variable (Max_PFN) which represents the total RAM addressed by the nucleus would artificially increase to 64TIB.
Since the GPU probably cannot access the entire 64tib range, it would signal dma_addressing_limited () as true. This variable essentially restricts the GPU to use the DMA32 zone, which offers only 4 GB of memory and explains performance regressions.
The good news is that this correction should be implemented as soon as the request for attraction lands, just before the Linux 6.15-RC1 The fusion window is closed today. With a general rate of six to eight weeks before the new Linux nuclei, we can expect the stable version 6.15 to be available to the end of May or early June.