David Vernet has proposed new AMD P-State driver logic that could reduce frame-time spikes in Linux games by temporarily boosting recently busy CPU cores. In his Steam Deck test, the change improved Civilization VI’s 1%-low frame rate by 31.8% and its p99 frame time by 4.1%.
The figures come from Vernet’s own testing in the Linux Kernel Mailing List patch series and have not yet been independently reproduced. The patch remains under review, so the result is promising evidence from one game rather than a general performance guarantee for AMD-powered Linux systems.
What does the AMD P-State patch change?
The proposed epp_boost mechanism adds an opt-in, per-core performance boost to the AMD P-State driver’s existing Energy Performance Preference support.
In active EPP mode, AMD hardware selects an operating point between configured minimum and maximum performance levels, guided by an EPP hint. Vernet says this approach can struggle when a game’s main or render thread repeatedly completes a burst of work and then sleeps briefly while waiting on a futex or GPU fence.
Each short sleep can weaken the hardware’s performance signal. The CPU may consequently resume at a lower operating point, increasing tail latency even though the core remains busy whenever work is available.
With epp_boost enabled, the driver samples each core’s C0 residency at most once every 10 milliseconds. If a core is at least 50% busy, its EPP value is temporarily set to performance mode. The normal cached request returns after 300 milliseconds without another busy sample.
The per-core approach is intended to avoid forcing every CPU core into performance mode. That is particularly relevant to handhelds such as the Steam Deck, where the CPU and GPU share a power budget.
How much smoother was Civilization VI?
Vernet’s Steam Deck benchmark recorded 31.8% higher 1%-low FPS and a 4.1% improvement in p99 frame time with epp_boost active.
| Metric | Reported improvement |
|---|---|
| 1%-low FPS | 31.8% |
| p99 frame time | 4.1% |
These measurements focus on frame consistency rather than average frame rate. Higher 1%-low FPS and better p99 frame times can mean fewer noticeable stutters and less severe frame-time spikes.
Phoronix reported the benchmark results from Vernet’s work, while the AMD P-State patch series on the Linux Kernel Mailing List provides the technical implementation and test details.
Civilization VI was selected for its reproducible benchmark, but no results were supplied for other games, AMD Ryzen handhelds, laptops, or desktop PCs. Wider testing will be needed to establish how consistently the patch helps across different processors and workloads.
Can Linux gamers enable epp_boost now?
Linux gamers can test epp_boost only by using a kernel built with the proposed patches because the feature has not been merged into the mainline Linux kernel.
On a patched kernel, the mechanism is enabled with amd_pstate.epp_boost=1. Its status can also be checked or changed at runtime through /sys/module/amd_pstate/parameters/epp_boost. If the parameter is not enabled, the driver continues using its previous behaviour without the new per-core boost logic.
The series still requires review, including scrutiny from AMD’s Linux kernel engineers. No Linux release has been named for possible inclusion, and the code may change during review.
For Steam Deck and AMD PC users, the practical decision is therefore to wait unless they already build and test custom kernels. The benchmark gives kernel developers a credible reason to examine the approach, but broader game testing will determine whether the gain extends beyond its carefully documented first case.
Leave a Reply