watchthelight labs

nvgfx: 2026-10-30 (full run, window 2026-10-29 10:04 UTC to 2026-10-30)

NullMoth shipped v2.6.4 (translator work only). The window drag thread got its first stock 2.6.3 number.

Actionable

#356, drag rate on stock 2.6.3. karma1228 (RTX 3060 Laptop, 1920x1080 165 Hz) counted 508 and 384 flips in two 8 s drags of a large window: 63.5 and 48 per second, with the GPU at 8 to 12 percent. Our 145 to 148 per second came from a scripted 800x500 Finder drag on our fork of 2.4.3, so the two are not directly comparable. Their new issue #456 puts WindowServer at 13 to 14 ms of CPU per presented frame, with the time in descriptor set allocation and update, the same profile as in #356. High confidence on the numbers, Medium on the cause.

#455, scaled HiDPI modes still missing on 2.6.4. Sequoia 15.8.1, RTX 4060 Laptop, 2560x1600 240 Hz. The log validates the 1680x1050 scaled mode, but Displays does not list it. Our fix for #354 went out in 2.6.3 (PR #421) and was tested on Tahoe with a 12-byte override. Whether this is a Sequoia gap or a different override format is unverified.

Notable

  • v2.6.4: texture arrays, sampler rewrites and a WebGL shadow fix in the translator. Nothing in the drag, hold or HiDPI paths.
  • #454: on an RTX 5080, WindowServer spins in NVRM when NVKMS rejects a modeset during a reconnect of an external DSC monitor. Our single eDP panel is unlikely to hit it.
  • #429: a 300 Hz eDP DSC panel goes black. Our panel is 165 Hz without DSC.
  • New related project: zzzyy678/SMCNvidiaGPU (MIT), a VirtualSMC plugin that reads NVIDIA temperature, power, fans and clocks on NullMoth 2.6.x through a table of NVRM offsets per build. It overlaps our nvgfx-smi; our builds would need their own table.

Noise

nullmoth/1401 #126, #124, #85, #123 and #26 (Meteor Lake, AMD and desktop installer reports); #390, #405, #328, #372; steelbrain/reims-vgpu (a virtual GPU for QEMU guests, GPL); e5u/FA608FM (AMD EFI); AgentBridge and rawtonic (unrelated).