Hello @cuda10358, we're using one issue report per unofficially supported game title. Usually I'd transfer this to #9789, but it doesn't appear to add anything of substance beyond what's already been written there.
OK. Cool. I just wanted to provide the info in case it did help with the ongoing issues. Thanks!
RADV_EXPERIMENTAL=heapx1 2026-06
I did use AI to aggregate some information for the report. I have verified its contents and added relevant information. The intent is to provide potential useful information for future optimization regarding how Mesa and the schedulers use AMD hardware to solve frame lag when loading world chunks. Please let me know if there's anything else I can do to contribute. I'm fairly versed in troubleshooting hardware and software. Lifetime Gamer and IT guy. XD
Thank you!
(Bug Report: Persistent Rhythmic Frametime Hitching (50ms->35ms) on RDNA2
System Information
Distribution: CachyOS (Arch-based)
Kernel: linux-cachyos-lts (6.x)
CPU: AMD Ryzen 7 5800X3D (Stock, PBO/CO disabled, Resizable BAR disabled)
GPU: AMD Radeon RX 6900XT
Driver: Mesa 24.x (RADV)
Proton Version: proton-cachyos-11.0-20260520-slr-x86_64
Issue Description
Persistent, predictable 35-50ms frametime spikes occurring while driving in the open world, regardless of graphics settings (High/Medium tested). This is not a hardware performance limit (GPU usage < 90%); it appears to be an engine-level asset streaming pipeline stall specific to RDNA2/Mesa interaction.
Troubleshooting Steps Taken (Variables Eliminated)
Storage: Verified issue occurs on NVMe Gen4 SSD (Bandwidth not the bottleneck).
Hardware Stability: Disabled PBO/CO, Reverted to stock BIOS settings, Disabled APE.
Kernel/Scheduler: Tested default bore kernel and lts kernel to rule out thread migration issues on 3D V-Cache.
Driver Overrides: Tested RADV_EXPERIMENTAL=heap and default configurations.
Controller: Using dedicated 2.4GHz wireless dongle (Bluetooth polling issue ruled out). Also used cable.
Desktop: Compositor suspended via qdbus to rule out X11/Wayland window management overhead.
Reproduction
Launch Forza Horizon 6.
Drive at high speeds (150mph+) in open-world segments.
Observe predictable frametime spike (50ms reduced to 35ms via this process) coinciding with world chunk loading.
Issue persists even when capping FPS to 90 via MangoHud.
Hypothesis
The issue appears to be a conflict between the Forza Horizon 6 asset streaming pipeline and the vkd3d memory descriptor heap boundary on RDNA2 hardware. The engine is waiting for the GPU to acknowledge the buffer swap during asset streaming.
Request
Could this be investigated for a potential vkd3d-proton patch or a specific workaround flag for the FH6 profile?)