Blog
What we shipped, and what it cost.
Release notes and engineering write-ups — including the bugs, which are usually the part worth reading.
- Engineering
Six networks, one all-reduce
An AI cluster is six different networks stacked. One 1 MB all-reduce, followed through all of them on one 36-core box, with the numbers each tier gave and the one each tier could not.
Read it → - Engineering
931 nodes, and where it broke at 3,267
A fourteen-pod fat tree with a working EVPN overlay on one box, the six bugs between here and there, and the honest account of the run that did not finish.
Read it → - Release
Labtris installs with one compose file now
A container install that does not touch the host's network stack, and the four bugs that only showed up once we ran a real lab in it.
Read it →