Measure X11-forwarded GUI latency headlessly: Xvfb -fbdir + memcmp, ssh ProxyCommand byte relay

To benchmark a remote X11 GUI (e.g. an EDA waveform viewer) over ssh -X without showing any window: run Xvfb :N -fbdir DIR locally (the framebuffer becomes an mmap-able XWD file), set DISPLAY=:N for the ssh client, and use a tiny C loop that memcmps the 6.4 MB frame against the previous one every 1 ms (0.15-0.5 ms per compare) and timestamps every change. Latency = last change before about 1 s of quiet. The 30 ms-stable rule undercounts badly, because over X the app often paints a small cursor/ruler update first and the main image 50-300 ms later. Count bytes per action with ssh -o ProxyCommand="python3 relay.py %h %p log" (a relay that timestamps every read each way). Drive the app through a separate ssh control channel so commands are not counted. Finding: a Qt-based viewer pushed about 5.2 MB per redraw (a full-window image) regardless of content. ssh Compression=yes cut that to about 30 KB and the latency from about 340 ms to about 60-100 ms on a 4 ms RTT link, so always check Compression when comparing X forwarding against anything. For a same-machine baseline, run a second Xvfb -fbdir on the remote host. Synopsys Verdi can be scripted for this with verdi -play file.tcl, where file.tcl installs a Tcl after-loop that polls a command file and logs clock microseconds around each eval.