Intel’s Graphics Driver Now Sharing ~60% Codebase Between Windows/Linux, 90~100% The Performance

Intel today announced the release of a server-based GPU for the Xe-LP data center, initially focused on high-density, low-latency Android cloud gaming and media streaming. For exciting things like Intel server GPUs, some interesting details about Intel Linux graphics drivers have also been revealed.

The GPU in the H3C server is your first discrete graphics processor for the data center. This first dGPU product based on the Gen12 Xe-LP has a 128-bit pipeline and 8GB LPDDR4 memory. The H3C includes four Intel Server PCI Express 3.0 x16 graphics processing units (GPUs) on a three-quarter long, high PCI Express 3.0 x16 board and supports over 100 Android cloud gaming users. Target configuration – up to four Intel H3C XG310 cards per server or sixteen Intel server GPUs per server. The GPU of the Intel server is sent, but price and availability information is still limited.

Before moving on to the exciting aspects of Linux, Intel also announced today that OneAPI Gold will be released in December and will bring the Intel Implicit SPMD Program Compiler (ISPC) to OneAPI level zero.

As Director of Graphics Software at Intel, Lisa Pearce introduced the Linux graphics drivers in her briefings and prepared them for the data center. She noted that two years ago only about 10% of the driver code was reused on different platforms (i.e. Windows and Linux) and that drivers for Linux provided about 60% of the performance of Windows. Moreover, they officially supported only one distribution.

But as we approach 2021, there is about ~60% code reuse, and Linux performance is 90~100%, which is also the case for Windows and has been tested on the three major Linux business distributions. The relative performance of 3D depends for about 90% on the quality of the game port of Linux or, if it is native, among other things, while the computing power is about the same between Windows and Linux.

Pending the emergence of Intel GPUs in other data centers, Intel is also introducing the GPGPU document portal. On dgpu-docs.intel.com there is a new portal with information about installing new drivers or updates for different distributions, the latest drivers and much more. Additional resources should be available in the area of GPGPU documents in the coming months.

Flipfast Project

The Project Flipfast initiative announced today is a new attempt to improve the gaming experience on Linux. The Flipfast project is designed to run graphics applications/games in a virtual machine, while the use of copy-free video memory buffers between guest and host ensures natural performance. This failover technology is also used by Intel designers in the data center to perform tasks in virtual machines.

Linux High for Intelgraphics

After that, I was able to talk to Lisa about the current development of the company with regard to the development of open source/Linux drivers. Of the ~60% of driver code shared by platforms, about 60% is represented by the Intel Graphics Compiler (IGC) and media/computer components currently shared by Windows and Linux. OpenGL and Vulcan drivers and kernel mode drivers remain separate codebases. This recent push to increase the development of Linux drivers is entirely due to their ambitions for data centers. Since ~2018, when the GPU plans for the data center were made, they started integrating their previously separated Linux graphics driver team and making other fundamental changes to support Linux.

CPI to Mesa

A few months ago I reported that an Intel Windows driver had started using PGI, and that there was even an internal prototype for Mesa. The conversation with Lisa, that the IGC is working for Mesa, is progressing. By modifying their NVA volcano driver to use IGC instead of the internal Intel Mesa compiler, they get promising performance results that seem to be much faster than the current compiler. Intel is also considering using IGC for its OpenGL Iris Gallium3D driver, but the Vulcan driver is very close for obvious reasons. Today, under Linux, the Intel graphical compiler is only used in the OpenCL/Level Zero Computing stack.

Although they offer better performance and more reuse of code on different platforms, there are still consequences, but for Intel Mesa drivers using IGC : Intel’s graphics compiler will be a new dependency for Mesa distributions, and Linux will have to package them, which will make the picture even bigger. The Intel IGC kernel binary comes with an additional 30 MB for extrusion in an ISO for Linux desktops.

They have the ambition to use the IGC in Intel Mesa code and will probably offer it in the first half of 2021, at least as a runtime alternative to their current shader compiler. This allows further testing of the PGI path and gives distributors more time to adjust their mesa-assemblies around the PGI and compare performance. If all goes well by the end of next year, we can conclude that the Mesa Intergovernmental Conference will be used by default.

GUI control panel for Linux ?

Also note that the radar with the Intel Linux graphics driver could have the Intel Control Center / Graphics Driver Control Panel by 2021. At this point the Intel Linux driver stack already contains most of the settings found in the Intel Windows driver area, but not in the GUI area. Most customizable and monitoring interfaces for the Intel Linux driver are only available via command line interfaces, installation of environment variables or module parameters that are not always well documented, etc. Lisa and her team want to offer something like a GUI control center for Linux systems, but not necessarily as complex as their Windows driver. They want more feedback on what Intel Linux GPU users would like to see in such a control center, and if we are lucky we might see it implemented in 2021.

These are the main conclusions of the Linux discussion. Intel’s efforts to open OneAPI’s open source software through the stack of graphics / computer drivers remains on a very exciting path, delivering good and reliable performance with near-perfect execution. On the Linux driver/software side there is something more to ask for because they basically already have their basics as they prepare to go further in the discrete graphics market. In case you missed it a couple of weeks ago, check out the Intel Xe graphics tests of OpenCL on OneAPI Level Zero and Vulkan.

If you liked this article, remember to join Phoronix Premium to see this site without advertising, multi-page articles on one page and other benefits. Tips from PayPal are also accepted. Thank you for your support.

Related Tags: