LibreBMC announced, end of Kestrel?
All PowerNV systems currently use some form of ASPEED BMC to provide the baseboard management features necessary to run the system (what we old fogies used to call "service processors"), which on these machines includes IPMI, offering the PNOR flash ROM to the processors to start the machine, a 2D frame buffer, environmental monitoring, network, front panel and other on-board interconnects. They run their own operating system, almost always OpenBMC, and function as a computer-within-a-computer. The goal here is to replace the ASPEED devices with a multi-vendor supported option compliant to the Open Compute Project's Data Center Secure Module Specification, but would (from our perspective as users and operators) appear much the same. It would also be compatible with Power, x86 and ARM systems like current BMC offerings. However, unlike ASPEED chips which are ARM-based, LibreBMC would run on a Power ISA core, possibly even a descendant of Microwatt.
Some of the names announced in connection with LibreBMC are familiar: Yadro appeared in yesterday's piece, a Russian developer of high performance systems and storage solutions, and Google was one of the developers of the OCP DC-SCM spec and is undoubtedly advising on the high-level design. Raptor is also involved, which is very good news, because their involvement suggests there are unlikely to be hidden blobs and the LibreBMC will be truly an open device (the announcement says it's all open tooling using LiteX, and synthesizeable on Lattice and Xilinx FPGAs). If you want a "little Power" system-on-a-chip, you could do a whole lot worse than simply ripping this off into a little stand-alone developer board. Maybe this is where the PowerPi will come from.
But, oddly, despite their involvement Raptor isn't making LibreBMC: that's being done by Antmicro, who currently has RISC-V and ARM-based products and will be "developing the LibreBMC card," implying that LibreBMC will also be offered as a physical component instead of just IP and VHDL code. Specifications for the physical form factor are even in the DC-SCM specification: it has a special slot similarly sized to an OCP NIC, but differently keyed, and the vertical form factor option 1 in particular looks very much like Beagle Boards and RPis. The fact that LibreBMC will run OpenBMC also suggests this is not a direct port of Kestrel, Raptor's own prototype BMC replacement; Kestrel currently runs its own OS based on Zephyr.
If this is a quiet way of announcing Kestrel is throwing in with LibreBMC, this would disappoint me somewhat. Kestrel, though granted in a very early form, seemed a better fit for those of us on the small but visible Power workstation side if its promises on faster start times were to be believed. In particular its trimmer OS choice was an important difference: OpenBMC's problem (from my outsider view) is that it wants to be all things to all workloads, which means a bigger loadout with more features but slower start times and a much larger firmware footprint. LibreBMC's move to a higher-performance Power core should help but software always expands to meet the available CPU power. BangBMC is still out there and aimed to address performance by cutting out fat like dbus and systemd, and apparently is now facilitated by Raptor, but it hasn't gotten a great deal of traction and nobody ships it as default, for example. A bespoke fast-start workstation-oriented BMC would be a great fit for our desktop systems but LibreBMC doesn't sound like that. And if this is going to be the simultaneous basis of a little Power board as well, bootup time really has to be better.
There are also the practical aspects for current owners: OpenBMC isn't going to drop ASPEED support anytime soon since there's a large install base running it (on lots of machines, not just OpenPOWER), but future updates would understandably prioritize the new hotness. Kestrel, although you needed a soldering iron, could be retrofitted to existing boards. Especially since open POWER10 systems will be slower to arrive, what about a LibreBMC retrofit option (that doesn't involve board work) for existing machines? Will the Antmicro "card" offer this, or will there be an aftermarket product from Raptor or another vendor? We'd also like whatever hardware and software improvements the LibreBMC initiative will foster especially if we won't be replacing our systems for awhile.
Nevertheless, I wait to see what happens with cautious enthusiasm, because I think that the importance of the BMC to the system (it runs and sees everything, after all) really demands an open solution. Plus, the presence of a BMC means there's (going to be) an OpenPOWER SoC, meaning smaller Power solutions may be just around the corner. Whoever ends up making it, this is good news.
There are already efforts to port BangBMC to kestrel.
ReplyDeleteHowever due to kestrels current lack of processing resources, its unlikely Raptor would ever ship it, or people would use it as a regular BMC.
Though there have been rumors that another POWER BMC system was in the works, and the BangBMC/Kestrel work would end up being useful there. Perhaps that rumor refers to the LibreBMC, I am not sure.