The real-time Linux kernel (PREEMPT_RT) is notoriously demanding: every component must ensure that critical sections do not block execution indefinitely. And if there is one component that has always been a thorn in its side, it is the Intel i915 graphics driver, the software backbone that manages Intel's integrated GPUs. After years of partial adaptations, a new series of patches is set to bridge the gap: the i915 driver is almost ready to coexist with the RT kernel, but under one precise condition: no display support. A limitation that, for certain scenarios, might be irrelevant or even an advantage.

To understand why this news deserves attention, one must start from the basics of real-time Linux. Unlike the standard kernel, where preemption can be disabled for extended periods by drivers or I/O operations, the PREEMPT_RT kernel minimizes non-preemptible regions, guaranteeing maximum latencies in the order of tens of microseconds. This is essential in fields such as robotics, industrial automation, critical infrastructure, and increasingly, AI model inference at the edge. Here, a delay of a few milliseconds in an accelerator's response can translate into an out-of-control movement or a missed safety decision.

The problem with GPU drivers is historical: GPUs have complex pipelines, command queues, and synchronization mechanisms between CPU and graphics engine that, in practice, introduce points where the kernel cannot be interrupted for potentially long periods. The i915 code, which covers a wide range of Intel chipsets, from Atom to the latest Core generations, is particularly intricate because it must also handle display functionality, mode setting, and frame buffer management. All of these are operations that, in a real-time context, become enemies of predictability.

The patches under review on the kernel mailing list tackle this difficulty at its root: they enable full compatibility with PREEMPT_RT only when the driver is configured to operate in "headless" mode, i.e., without any video output. Technically, this means that display management routines are excluded from the build, removing the primary source of unpredictable latencies. The graphics chip remains accessible as a compute engine: through APIs like OpenCL, oneAPI Level Zero, or inference libraries such as OpenVINO, it can be exploited for parallel computation. It is no coincidence that the patch has been welcomed with interest by those working on AI workloads in industrial settings: Intel integrated GPUs are ubiquitous in embedded systems and compact servers, and making them deterministic would open the door to on-premise deployments where dedicated hardware (such as expensive NVIDIA GPUs) is not justifiable on cost or power grounds.

For those evaluating self-hosted inference solutions with LLMs or computer vision models, the trade-off is evident. On one hand, the ability to connect a monitor is lost, which is not a problem in an edge node mounted in a factory cabinet or in a headless server. On the other hand, there is the certainty that model processing will not suffer from jitter introduced by the operating system. The TCO (Total Cost of Ownership) of a fleet of machines based on Intel processors with integrated GPUs could be significantly reduced, eliminating the need for external accelerators for workloads that do not require massive computing power but demand responses within tight time windows.

It should be noted that the situation is still in flux: the patches must pass the final round of reviews and convince the maintainers of the Direct Rendering Manager graphics subsystem. However, the growing consensus suggests that inclusion in the mainline kernel is only a matter of time. Intel, for its part, has invested heavily in recent years in Linux support, both for CPUs and GPUs, and this move fits into a broader strategy of enabling open platforms for edge computing. It remains to be seen whether, with the arrival of dedicated Arc cards and the evolution of the Xe driver, the same RT philosophy will be extended to the new stack, perhaps with fewer compromises.