64 bit ARM systems with more than four cores

[greg at unrelenting.technology]

December 2, 2019 12:37 AM, "John F Carr" <jfc at mit.edu> wrote:

> Two RK3399 based systems, 2 A-72 + 4 A-53: RockPro64 (http://www.pine64.com) and ROCK Pi 4
> (http://radxa.com). I can't use the RockPro64 due to the 1.5 MBps serial port. Apparently it works
> or almost works with some special handling if your serial interface can count to 1,500,000. The
> ROCK Pi 4 is in crochet so it must work... right? Does it have a reasonable console port bit rate?

Before I got my 1.5M capable dongle for the ROCKPro64, I just had an u-boot command to
switch to 115200 in the boot script.

I consider 2xA72+4xA53 worse than 4xA72 to be honest.
Especially since the FreeBSD scheduler is not big.LITTLE aware — it loves to schedule
e.g. a big single-threaded linker job onto the slow cores. So you have to mess with
cpuset but apply it carefully to still utilize all cores on parallel tasks.. argh.

> One LX2160A based system, 16 A-72 cores: HoneyComb LX2K
> (https://www.solid-run.com/nxp-lx2160a-family/honeycomb-workstation). I see relevant files in
> sys/gnu/dts/arm64/freescale but nothing outside of these files imported from Linux. It's not
> mentioned on https://wiki.freebsd.org/arm64. My guess is that means FreeBSD does not run. Is it a
> little job or a big one?

There's SBSA-ish ACPI-capable firmware so it should boot, but probably with uhh not a lot of devices.
They've been sort of vaguely promising that PCIe would work in ACPI..
and recently this commit dropped:
https://source.codeaurora.org/external/qoriq/qoriq-components/edk2-platforms/commit/?h=LX2-UEFI-ACPI-0.2&id=ec8deb524171a03448395e6b8acc49978bcf5ce5
so maybe I shouldn't have doubted them, but I missed the $500 discount period waiting for this :(

> SC2A11, 24 A-53 cores at 1 GHz: SynQuacer (https://www.96boards.org/product/developerbox). I don't
> see any evidence of SC2A11 support in the kernel tree or on https://wiki.freebsd.org/arm64. My
> guess is that means FreeBSD does not run.

There's SBSA-ish ACPI-capable firmware so it should boot, and PCIe should work.
I guess nobody tried because it's very expensive for something with A53 cores.

I would *not* trade my MACCHIATObin for this because A53 is just so slow.

> ThunderX in various forms, rack mount or workstation. Nice specs and apparently supported but the
> two American system builders don't seem interested in selling me one.

Keep in mind that the TX1 cores are slow. It's great at super parallel workloads,
but when you have something single-threaded, it's no better than an A53.



Now, if you have cash but not ThunderX2 levels of cash: Ampere eMAG is the best option.
Available in Lenovo HR350A and HR330A boxes, also from Avantek (idk what's the mainboard there??),
also rentable in bare-metal-cloud at Packet (they have the Lenovos).

See https://bugs.freebsd.org/bugzilla/show_bug.cgi?id=237055 for the bringup story :)
I think most of the bugfixes required for eMAG are in -CURRENT.