nvgfx: 2026-10-27 (full run, window 2026-10-23 23:28 to 2026-10-27 09:30 UTC)

Our drag fix, measured on an RTX 5080, and one regression in it. nullmoth/nvidia-macos-driver#412: a user took six of our NvGfx plugin commits, rebased them onto NullMoth 2.6.2 and measured window drags on an RTX 5080. Stock 2.6.2 gave 85 to 100 flips/s with stalls of 63 to 124 ms; with our commits the stalls go away. That is the first measurement of the I-069 fix on hardware other than ours. They held back two commits, the ones that cache setRenderPipelineState and the sampler variant, because with them Alacritty (OpenGL) text jumped backward. Two more did not apply cleanly (the System Settings icon speedup and the Chrome CPU-raster scroll). We need to check the two state-cache commits against an OpenGL terminal before we call them done. High confidence: their own numbers, our code.

Brightness that survives a restart. PR #312 keeps a logical brightness target from 0 to 65535 across power requests and maps it to 5 to 100 percent on the panel. The submitter now reports the saved level comes back after a restart on an RTX 5080 Laptop under macOS 26.7. Still open and unreviewed. NullMoth has no licence, so we would rewrite the idea, not copy it. Medium confidence (one machine).

HDR is blocked by the class name, not the pixel format. In #341 a user traced CoreDisplay's vendor check: a framebuffer class whose name contains "NVDA" is treated as vendor 8 and gets force8BPC = YES, so 10-bit and HDR timings are never considered. Advertising 10 bpc, BT.2020 and HDR10 did not change SupportsHDRMode. Renaming the class cleared the flag but WindowServer then lost its Metal compositor and aborted. This explains why our 30-bit attempts (N-187 to N-190) could not get further. Medium confidence, one report, not reproduced by us.

HiDPI: the missing flag and the override format. In #354 the fix we pointed at was confirmed: getInformationForDisplayMode() sets only Valid, Safe and NotPreset, and scaled modes need kDisplayModeValidForHiResFlag (0x00800000) to reach the public list. In #341 and #360, independent reports on GA106 confirm that 12-byte scale-resolutions entries with backing sizes work, 16-byte entries carrying flag 0x00200000 hide the modes, and scales above about 1.3x at 165 Hz hit a scaler bandwidth limit. A desktop RTX 3060 at 120 Hz over DisplayPort shows the same refusal. Matches our I-059 notes. Medium confidence, not merged upstream.

The shader cache cost is a cold-cache cost. The reporter of #370 measured the translator's cache and found the writeToFile/fsync storm only happens while the cache fills after a driver update. So it comes back after every update. That supports moving the write off the hot path, which is what our cache-write-async branch does. High confidence on the measurement.

Releases 2.5.5 to 2.6.2. 2.5.5 fixes Discord video, encoder keyframes, Screen Sharing (buffer-backed textures accept any view format) and ANGLE Metal in Brave and Edge. 2.6.0 makes a driver update need one restart; 2.6.1 and 2.6.2 are the same driver with installer fixes. Read from the release notes and commit titles only.

A frame missed every 0.6 s. #410 reports every app missing a frame at a steady 0.6 s rhythm with WindowServer blocked in the kernel at the same rhythm. A fix PR is planned. Worth watching next to our drag work. Unverified.

mickys/nvidia-macos-driver (2.5.4-mickys.1): thirteen games/* branches with translator fixes (phi on local fields, gather-compare, cube-as-array, struct-member arrays), presenting drawables on a serial queue, and 70 Steam games tested on a 4090. The present-queue branch sits close to our drag work.

nullmoth/1401 2.5.5 to 2.6.2 (installer and EFI checker), nullmoth/1401-probe (privacy fix), thanhnndev's fork (an AMD/Intel NMXAccel idea), steelbrain/reims-vgpu (a virtual GPU for guests), Jakiewbe/honor-hackintosh-lab, MegaMind2566's IdeaPad S540 EFI, #333 (Netflix Widevine, not graphics), #414, #403. Discovery: zpalmtree/seine, AnalogicGoose/tuf-f15-controls, leyodr9000/firebat-t5a-hackintosh, moeru-ai/airi, radiumcoders/localflow.

• nullmoth#410: every app misses a frame about every 0.6 s (RTX 5060 Ti desktop, Tahoe 26.7.1, 2.6.2). The reporter's WindowServer spindump shows it blocked in the kernel at the same rhythm and says a PR is coming. Our flip-counter traces can say whether we have the same periodic gap, and the PR may overlap our drag work. Medium confidence.
• nullmoth#419: a translator patch from eisda against 2.6.2, reported to take translation failures from 53 to 0 across 349 Chrome modules, and a demo to 64.5 fps over 18,066 frames. The translator is LGPL-3.0-or-later (translator/translator/Cargo.toml in our tree), so the patch is reusable if we keep the licence. No maintainer reply. Untested by us. Medium confidence.

• No 2.6.3. 2.6.2 (26 Oct) is the newest; 2.6.0 makes updates need one restart, and 2.6.1 and 2.6.2 only change the 1401 apps. High confidence.
• #406: an RTX 5050 Laptop (GB207, 10de:2d98) on Tahoe 26.7.1 fails Unity WebGL with the translator error "native emitter: owned IAdd operands do not match its result type". Same family as #419. Another RTX 50 laptop on Tahoe.
• PR #418: the AIR parser read !dbg metadata as operands, which broke MSAA shaders (GTX 1660 Ti, DaVinci Resolve). Different failure; nothing for us unless we hit it.
• #412: TroyMoss16 cut his set to four patches. Live corner resize still stutters on his RTX 5080 (95 to 108 flips per second). No maintainer reply.
• HiDPI (#341, #360) and HDR (still 8 bit, #30) have nothing new. PRs #408 and #409 (BAR1 placement without a 64-bit _CRS window, RTX 2060 on an X399 board) are desktop-only.
• nullmoth/1401 boot hangs after EXITBS:START are covered in legion-efi.

• KangweiZhu/ROG-G16-2024-Hackintosh-QEMU-KVM: a macOS 15.8.1 VM with an RTX 4060 Laptop passed through, on NullMoth 2.5.3 with BAR1 cut to 256 MiB. MIT scripts, four commits, no stars. Related, marginal.
• joseteo/nvidia-macos-opencore (a Rust CLI for Turing laptops on Sequoia, MIT) is about 15 days old, outside the window. Noise.
• Press and SEO pages that say "Sequoia only, no Tahoe" contradict the NullMoth README and our own boots. Noise.

Jakiewbe/honor-hackintosh-lab (five CI commits), steelbrain/reims-vgpu (one install issue, no commits), alexandreunitypro/MetalHorizon-FH6-Mac-Launcher, FFGG10-29/Y7000-2018-Hackintosh, Nanigummadi/ASUS-TUF-F15-FX506HC-Hackintosh-Mac-OS, power418/hp-victus-15-hackintosh.