Banana Pi to Launch a 24-Core Arm Server

SinoVoIP has been offering Banana Pi single board computers for several years. Their boards are generally based on Arm processors, offered at a relatively decent price, although not as quite as good value as FriendlyELEC and Orange Pi ones. The company is also known for providing subpar documentation and firmware images, but a fairly active community still formed around their boards 🙂

The company has now demonstrated something a little different with a 24-core Arm server that should eventually be sold as a Banana Pi server board or actual server, as the full details are yet to be known.

Banana Pi 24-Core Arm Server
Click to Enlarge

We did not get a glimpse at the actual hardware, but the blurry photo above gives some clues. We have 24-core Arm Cortex A53 processor with 32GB RAM (29.4GB seen by the OS) running Ubuntu 18.04.1 LTS with MATE desktop. There aren’t that many 24-core Arm Cortex A53 processors, so unless the company is using an announced processor, it has to be SocioNext SC2A11 processor also found in Linaro Developer Box.

The video below shows the server’s 24 cores fully utilized while building Linux 4.19 kernel.

Nora, Project Manager for Banana Pi at Foxconn, provides a few more details in LinkedIn comments and we know the board supports NVMe storage, the company tested TensorFlow under Docker, Raspbian, and ROS Melodic Morenia. We’ll likely have to wait a few more months to find out details about hardware, price and availability.

Share this:
FacebookTwitterHacker NewsSlashdotRedditLinkedInPinterestFlipboardMeWeLineEmailShare

Support CNX Software! Donate via cryptocurrencies, become a Patron on Patreon, or purchase goods on Amazon or Aliexpress

ROCK Pi 4C Plus

26 Replies to “Banana Pi to Launch a 24-Core Arm Server”

      1. I sincerely doubt this would be a server part in the sense most of us imply. More like a many-core SBC with a proper DIMM slot or two.

    1. Makes no difference for highly parallelized tasks. It’s 24 cores, so it’ll be about as fast as a 6core cpu at 4Ghz in parallelized tasks (which is why you’d get it for).
      The soc needs to be re engineered though.
      1Ghz for power efficiency is old 28+nm technology. At below 14nm, the most optimal speeds are 2Ghz with boosts to 2.5Ghz.

      1. I brought up the clock stricktly for identification purposes. Speaking of which, do we have the lscpu from an actual SC2A11 — I’m curious if that one shows 8 nodes x 3 cores (the physical layout of the SoC), as this one shows 12 nodes x 2 cores..

  1. I’m going to get some popcorn and wait for tkaiser’s takedown of this whole concept.

    Quick, do you think SinoVoIP have the power states correct for the CPU? Previous experience would suggest no 😉

    1. people like you are the reason we dont have nice things, but crap. and getting worse every generation of clueless consumers.

  2. > as the title implies it’s running the recent Linux 4.19

    The video is titled ‘ubuntu 18.04 build kernel 4.19′ and their forum post shows a screenshot that suggests the ‘server’ is running 4.9.82 instead. But since this company is not able to provide information anyway why bother…

    If it’s SC2A11 then good luck calling this a ‘server’ with that limited I/O capabilities…

      1. Per core.
        According to documentation, one can add up to 64 codes on a board, that’s almost 1TB on ram, not shared by cores.
        Somehow this has to be a typo, 64 bit supports up to 1TB of ram.
        No one said anything external bridge chip couldn’t be used for extra Ram.

  3. LOL another “demo” from Sinovoip … the useful version of software ( plus documentation ) will be ready in 2028 …

    I am still waiting for a usable BPi W2 kernel .. Never waste time on boards from this crap outfit.

    1. Lets hope they chose a really badly supported soc without any mainlining at all 😉 will be fun for many early adopters 😀

  4. My idea. For real.
    Servers now are hopping on the Qualcomm centiq 2400, but for a fraction of the price, you can run your company on a banana pi.
    Ow, and Linux version doesn’t matter. You can install just about any Linux on it that supports ARM hardware. Including android.
    But really, 24 arm codes, they should at least double it. Make it a dual core board, 48 threads, running on a 19V laptop power supply.
    24 cores, you might find a Xeon 10 core machine for about the same price, and it would outperform it save for power requirements.

    1. If it is indeed SC2A11 processor, you can connect multiple processors over PCIe for a total of up to 1536 cores.

  5. Die hards will be saying the raspberry pi server SoC is due late 2019 with 100 cores and 6 10 gigabyte ethernet connected to the usb2 hub, 1GB ram. No heatsink required.

      1. > Raspberry has already proven itself useful in MPI many clusters

        Useful for educational purposes, sure, but most high-core servers are built for other purposes than just education.

        > I’d expect a 24 core server RPi in few years.

        Reading “server” and “RPi” juxtaposed in the same sentence makes me think “why?”. The *only* value of RPi is its community. Its reliability and performance are lower than what you can get from a lot of other less expensive hardware. People building servers want reliability and most often performance (especially when buying many cores). This is not in the DNA of RPi at all. RPi is about bringing small computers to kids to tech them the basics of computer science with something that more or less works even when heavily molested (dog chewing power cable, or improperly installed heat sink). It really doesn’t match what professionals are looking for. Maybe you’ll see high core counts just for the purpose of teaching parallelism in schools though, but noone serious would want to build a server based on this, and the limited ability to sell such a product will make it too expensive for any useful purpose.

  6. Really strange that the lscpu tool shows 12 sockets and 2 cores per socket instead of 24 cores on single socket.

    1. Clusters do tend to show as numa nodes in numa topologies, and this is normal. The thing I’ve bee trying to figure out, is given that the original 24-core SC2A11 has a physical topology of 8 clusters by 3 cores, not 12 clusters by 2 cores, as shown in the screenshot, what does an actual SocioNext machine report. Of course all that can be configured in various ways, and in the case of multiple-clusters per socket, it would not make much of a performance difference, given there’s still a single memory controller that services all the clusters.

Leave a Reply

Your email address will not be published. Required fields are marked *

Khadas VIM4 SBC
Khadas VIM4 SBC