Intel's Open-Source Linux Compute Stack Maturing Very Well For Arc Graphics
From early December to late February there was an absence of new Compute-Runtime updates for that open-source stack for providing OpenCL and oneAPI Level Zero support for Intel graphics hardware on Linux. It was out of trend as they worked to move from a weekly~biweekly release rhythm to a monthly release cadence while taking extra time for making various other changes too. After that three month lull, they are back to pushing out new compute updates and damn it's looking nice. At least in my testing, the progress they've quietly made over the past few months has been very nice for the compute stack compatibility/support and performance.
I've been running the Compute Runtime 22.53.25242.13 stack paired with Linux 6.2 for nearly two weeks now across my Intel Arc Graphics A380, A750, and A770 graphics cards and it's been working out very well. Superb. A night and day difference compared to the various other times trying it on older Linux software versions since the A750/A770 launch. In the past there were issues with random OpenCL programs, the Blender oneAPI back-end would seg fault for me when launching it via the command line while working fine from the GUI, and other oddities. But now for the past two weeks this latest Intel open-source compute stack has simply been running great with whatever compute workloads I've been throwing at it.
The maturity point of the Arc Graphics DG2/Alchemist compute driver also times well considering that it's with the recently released Linux 6.2 where the DG2/Alchemist graphics are enabled by default after long being behind a module parameter flag for the then "experimental" hardware support. It's also with Linux 6.2 where there is also HWMON power reporting for DG2 GPUs and other improvements.
Long story short, the open-source Linux driver stack for Intel Arc Graphics now is in great shape. Besides the open-source OpenGL and Vulkan drivers maturing the past few months, I've been very happy with how the latest Compute Runtime stack has been running and performing across the A380 / A750 / A770 consumer cards with all past issues/quirks seemingly cleared up.
Setting up the Compute Runtime on Ubuntu also remains a breeze -- even using the binaries on the yet-to-be-released Ubuntu 23.04 worked without issue and atop a vanilla Linux 6.2 kernel. (Compared to the often fragile Radeon ROCm stack when running on non-enterprise / non-LTS releases.) Plus Intel's GPU compute stack is fully open and built on open standards, in stark contrast to NVIDIA's CUDA ecosystem.
For seeing how the A380/A750/A770 are running on this latest compute stack paired with Linux 6.2, a variety of OpenCL and Blender benchmarks were run on those cards as well as the NVIDIA GeForce GTX 1660 SUPER / RTX 2060 / RTX 3060 graphics cards while using the proprietary 530.30.02 Linux driver stack. Then with the ROCm 5.4 compute stack on Linux 6.2 was then also testing the Radeon RX 6500 XT, RX 6600, and RX 6700 XT graphics cards. All testing happened on the AMD Ryzen 9 7950X system with ASUS ROG CROSSHAIR X670E HERO motherboard, 2 x 16GB DDR5-6000, WD_BLACK SN850X 1TB NVMe SSD, and the various graphics cards. Ubuntu 23.04 with the Linux 6.2.2 kernel was in use for all testing and that kernel build obtained from the Ubuntu Mainline Kernel PPA for easy reproducibility.