The YouTube video recordings for the X.Org Developers’ Conference 2025 that took place at the end of September in Austria are finally available. Among the many interesting XDC2025 presentations was Intel engineer Matthew Brost talking about the GPU Shared Virtual Memory (SVM) within Intel’s modern Xe kernel graphics driver.
Shared Virtual Memory (SVM) and Heterogeneous Memory Management (HMM) shouldn’t be anything new for longtime Phoronix readers or in regards to Intel their ongoing work in that area with earlier this year merging SVM for Xe and now also tackling multi-GPU support in Xe as part of their ongoing Project Battlematrix initiative.
Matthew Brost confirmed in his talk that GPU SVM in Xe is working nicely for single-device purposes, userptr “user pointer” support is alos in place, and other bits while the multi-device support is still a work-in-progress as are more performance optimizations.
Brost reaffirmed PCI Express peer-to-peer (PCIe P2P) plans as well as confirming the Intel Xe Linux driver also has plans for supporting high speed fabrics such as UALink.
While Intel as a company supports the Ultra Accelerator Link (UALink) standard, to date there hadn’t been any real direct mentions of UALink around the Xe kernel driver. UALink as a reminder is the cross-vendor specification as a die-to-die interconnect and serial bus between AI accelerators. Given the dominance of Linux in HPC and AI, not too surprising but nice seeing them officially mentioning their plans for supporting high speed fabrics in Xe like UALink. But no details beyond that confirmation were shared during XDC2025.
The XDC2025 presentation is embedded above and the Power Point slides from the presentation can also be found here.
