Repository navigation
🎛️ Every card in a multi-GPU box. A Linux machine with more than one NVIDIA card
now shows each of them in Fleet: a chip per card with its load, memory and temperature,
and the page below — charts, details, processes — describes the card you pick. Each
card keeps its own history, so switching cards doesn't mix their charts. The overview
tile charts up to two cards and lists the rest; the sidebar shows every card's load.
Processes on the card they actually use. The Linux agent matches each GPU process
to its card by the card's UUID, so a job on GPU 1 is listed under GPU 1, and a process
that uses two cards appears under both, with the memory it holds on each. Every
process also shows the account it runs as — on a shared box, that is who holds the
card. An older agent lists every process under the first card, as before.
🌐 Your local addresses. The Network section shows each interface's IPv4 and IPv6
addresses in its row; expand it for the full, copyable addresses. VPN and tunnel
interfaces are listed under More addresses. Nothing is looked up online.
(#79, @evn88)
Fixes — all four from @nova28:
- Fleet could crash the app while polling, when a connection was closed just as a
request was about to use it. (#83) - With notifications on and the window closed, the GPU throttle alert was judged on
the last temperature and power measured while the window was open. It now measures
them. (#81) - A menu-bar dropdown now gives back its memory when it closes. Before, it held on to
about 110 MB of graphics memory for the life of the app on a 5K display.
(#80) - A remote Mac whose GPU power is still pending, or that has no GPU sensor, no longer
reads "0W · 0°C" in Fleet; the value is left out.
(#82)
The Linux agent is now 1.5.0 for the per-card processes and accounts above; the Mac
agent is unchanged at 1.4.1.
