Facebook Yosemite v3.5 Platform and CraterLake Server (fby35)

Facebook has a series of multi-node compute server designs named Yosemite. The most recent version released was Yosemite v3.

Yosemite v3.5 is an iteration on this design, and is very similar: there’s a baseboard with a BMC, and 4 server slots. The new server board design termed “CraterLake” includes a Bridge IC (BIC), with room for expansion boards to include various compute accelerators (video, inferencing, etc). At the moment, only the first server slot’s BIC is included.

Yosemite v3.5 is itself a sled which fits into a 40U chassis, and 3 sleds can be fit into a chassis. See here for an example.

In this generation, the BMC is an AST2600 and each BIC is an AST1030. The BMC runs OpenBMC, and the BIC runs OpenBIC.

Firmware images can be retrieved from the Github releases or built from the source code, see the README’s for instructions on that. This image uses the “fby35” machine recipe from OpenBMC, and the “yv35-cl” target from OpenBIC. Some reference images can also be found here:

$ wget https://github.com/facebook/openbmc/releases/download/openbmc-e2294ff5d31d/fby35.mtd
$ wget https://github.com/peterdelevoryas/OpenBIC/releases/download/oby35-cl-2022.13.01/Y35BCL.elf

Since this machine has multiple SoC’s, each with their own serial console, the recommended way to run it is to allocate a pseudoterminal for each serial console and let the monitor use stdio. Also, starting in a paused state is useful because it allows you to attach to the pseudoterminals before the boot process starts.

$ qemu-system-arm -machine fby35 \
    -drive file=fby35.mtd,format=raw,if=mtd \
    -device loader,file=Y35BCL.elf,addr=0,cpu-num=2 \
    -serial pty -serial pty -serial mon:stdio \
    -display none -S
$ screen /dev/tty0 # In a separate TMUX pane, terminal window, etc.
$ screen /dev/tty1
$ (qemu) c             # Start the boot process once screen is setup.

This machine model supports emulation of the boot from the CE0 flash device by setting option execute-in-place. When using this option, the CPU fetches instructions to execute by reading CE0 and not from a preloaded ROM initialized at machine init time. As a result, execution will be slower.