Research Digest 0734: NVIDIA R610.43.02 Driver Changelog Audit¶

Date: 2026-05-28 Branch: docs/r610-driver-changelog-audit-20260528 Driver installed locally: 610.43.02 (loaded after reboot per user confirmation) Scope: Assess driver-level changes affecting CUDA workloads (UVM, scheduler, power, memory bandwidth); map findings to fork CUDA paths in core/src/cuda/ and cmd/vmafx-server/.

1. Changelog fetch status¶

Conclusion: NVIDIA has still not published a standalone changelog.html for the 610.43.02 series under the canonical XFree86 README path. The directory index at README/ is live but contains only the standard installation guide sections (Introduction through Appendix L); no changelog entry point exists.

A substantive forum thread ("610 release feedback & discussion") was found via the category index and yielded the content below. All quotes are sourced from that thread as of 2026-05-28.

2. Confirmed R610.43.02 changes (from developer forum thread)¶

2.1 Vulkan extensions added¶

"Added support for the following Vulkan extensions: VK_EXT_shader_long_vector VK_KHR_internally_synchronized_queues VK_NV_push_constant_bank"

"Added support for creating Vulkan logical devices from multiple physical devices on select cards via VK_KHR_device_group_creation. This feature can be enabled by setting the environment variable __VK_ENABLE_DEVICE_GROUPS=1."

2.2 Display and framebuffer¶

• Added support for FP16 EGL framebuffer configurations on Wayland.
• Added support for DRM format modifiers for multiplanar YCbCr formats.
• Added support for mmap on DMABUF file descriptors exported from discrete NVIDIA GPUs.
• Added support in the nvidia-drm kernel module for the per-plane DRM color pipeline API introduced in Linux v6.19. A new color_pipeline kernel module parameter was added to nvidia-drm to allow disabling this support as a workaround.

2.3 Removals¶

• Removed support for using the NVIDIA X11 driver with Xinerama.

2.4 Bug fixes vs R580 series¶

• Fixed a Vulkan performance regression (vs upstream R580 series) that affected at least one shipping title.
• Fixed regressions introduced in 580.65.06 and 580.105.08.

2.5 GSP firmware¶

GSP firmware path: /lib/firmware/nvidia/610.43.02/ (gsp_*.bin, architecture-split). Enabled by default on Turing and later. No new GSP-level changes documented.

3. CUDA / UVM / scheduler findings¶

None confirmed. The forum thread, directory README, and all reachable 610.43.02 documentation are silent on:

• UVM (Unified Memory) subsystem changes
• CUDA kernel scheduler changes
• Concurrent multi-process execution (MPS) changes
• Power management or GPU clock policy changes specific to compute workloads
• Memory bandwidth or NVLINK changes
• New driver-level environment variables affecting CUDA (no new NVIDIA_* or CUDA_* knobs documented)

The known-issues section of the README covers only display/graphics topics; no compute or UVM regressions are listed.

4. Impact mapping to fork code paths¶

4.1 core/src/cuda/picture_cuda.c — CUDA picture pool / pinned memory¶

No driver-level UVM or pinned-memory changes were found in R610.43.02. The async cuMemcpy2DAsync pattern used for upload/download in vmaf_cuda_picture_upload_async / vmaf_cuda_picture_download_async is unaffected.

Assessment: No action required.

4.2 core/src/feature/cuda/ — CUDA feature kernels / concurrent execution¶

No scheduler or concurrent-kernel-execution changes were found. The kernel concurrency behavior on the RTX 4090 (Ada Lovelace, Turing-lineage GSP) is unchanged by this driver drop.

Assessment: No action required.

4.3 cmd/vmafx-server/ — gRPC server (multi-request concurrency)¶

No MPS, CUDA context, or process-isolation changes were documented. The server's concurrency model (per-request CUDA context sharing via the fork's dispatch strategy in core/src/cuda/dispatch_strategy.c) is unaffected.

Assessment: No action required.

4.4 DMABUF / DRM format modifier changes¶

The addition of mmap support on DMABUF FDs exported from discrete NVIDIA GPUs (Added support for mmap on DMABUF file descriptors) is potentially relevant to the fork's SYCL USM/dmabuf import path (core/src/sycl/), not CUDA directly. This is a kernel-module improvement that widens what user-space can do with exported dmabuf handles. No code changes are required now; if a zero-copy import path is added for the CUDA backend in the future, the DMABUF mmap capability in R610 is a prerequisite that is now confirmed satisfied.

Assessment: Informational; no immediate action.

5. Summary verdict¶

R610.43.02 is a display-focused release relative to the R580 series. The changes that are confirmed (new Vulkan extensions, DRM color pipeline, FP16 EGL, DMABUF mmap, Xinerama removal) are all in the graphics/Wayland/DRM layer and have no effect on CUDA compute paths. The canonical changelog.html remains unpublished at NVIDIA's XFree86 index. No driver-level UVM, scheduler, power-management, or CUDA-specific changes were found that require code or configuration changes in the fork.

6. Reproducer¶

# Primary source (200 — live at time of research)
curl -sI 'https://download.nvidia.com/XFree86/Linux-x86_64/610.43.02/README/'

# Still 404 at time of research
curl -sI 'https://download.nvidia.com/XFree86/Linux-x86_64/610.43.02/README/changelog.html'

# Forum thread (200 — category index, thread "610 release feedback & discussion")
# https://forums.developer.nvidia.com/c/gpu-graphics/linux/148