Radeon Vega 20 Will Have XGMI - Linux Patches Posted For This High-Speed Interface

It has been expected that Vega 20 would feature XGMI as a high-speed GPU interconnect alternative to PCI Express and that was firmed up today thanks to a new set of AMDGPU Linux driver patches.

XGMI is a peer-to-peer high-speed interconnect and is based on Infinity Fabric. XGMI is basically AMD's alternative to NVIDIA's NVLink for inter-connecting GPUs. Previously leaked slides indicated that XGMI would be supported on Vega 20 along with PCI Express 4.0. That appears accurate as becoming public today were a set of patches wiring in XGMI support to the AMDGPU Direct Rendering Manager driver and explicitly enabling the functionality for Vega 20. It's also expected AMD's upcoming Naples server CPU architecture will also support XGMI.

This is the first time we're seeing XGMI patches for the AMDGPU Linux driver. They for now can be found on the amd-gfx list but will likely be queued via -next for introduction in the Linux 4.20~5.0 kernel as part of their latest Vega 20 enablement work.

The 13 patches out today don't appear to reveal any exciting details beyond confirming Vega 20 supporting this tech.

Overall, the Vega 20 open-source Linux driver support appears to be squaring up nicely in time for this expected compute/HPC 7nm Vega part that's expected to launch by the end of 2018. Details are still light but the open-source patches have confirmed it's a discrete card, adds new deep learning instructions, etc. With the next kernel cycle to be known as Linux 4.20 or 5.0, the Vega 20 driver support is no longer experimental but bits are still landing like this XGMI high-speed interface support.