Slashdot Mirror
Banana Pi 24-Core ARM Server Running Ubuntu Breaks Cover (hothardware.com)
MojoKid writes: ARM-based server processors have threatened to take on Intel in the data center for some time but not much has materialized thus far in terms of significant deployments. However, a new breed of low cost ARM server implementations may be in the works with a many-core platform called Banana Pi. The latest Banana Pi device being teased is something very different in the form of a 24-core ARM server that speculation suggests might be sold as a Banana Pi server board or as a finished server product.
A video has surfaced that reportedly shows a 24-core ARM Cortex-A53 processor with 32GB of RAM, though the OS only sees 29.4GB of that RAM. The OS is Ubuntu 18.04.1 LTS with MATE desktop. Unless the processor used in this device is something unannounced, and that seems unlikely, the chip itself would likely be a SocioNext SC2A11. The same processor is used in the Linaro Developer Box. The demo shows the server fully loaded at 100% CPU utilization building a Linux kernel and reportedly the system also supports NVMe storage as well as TensorFlow workloads for machine learning. Not much else is known about the system at this time but it's an interesting development in the Linux server space to be sure.
A video has surfaced that reportedly shows a 24-core ARM Cortex-A53 processor with 32GB of RAM, though the OS only sees 29.4GB of that RAM. The OS is Ubuntu 18.04.1 LTS with MATE desktop. Unless the processor used in this device is something unannounced, and that seems unlikely, the chip itself would likely be a SocioNext SC2A11. The same processor is used in the Linaro Developer Box. The demo shows the server fully loaded at 100% CPU utilization building a Linux kernel and reportedly the system also supports NVMe storage as well as TensorFlow workloads for machine learning. Not much else is known about the system at this time but it's an interesting development in the Linux server space to be sure.
have threatened to take on Intel in the data center for some time but not much has materialized
Counterpoint: since May 2016, a French hosting company called OVH has been offering ARM-based servers under the brand Scaleway. I've been happy with it, mainly because of the price: EUR. 2,99 per month (US$ 3,41). I mainly use it to host my personal websites, a self-hosted TT-RSS instance, a private wiki, that sort of stuff. You get an ARM board with four cores and 2 gigs of memory. And no worrying about the latest Intel problems like Spectre, Meltdown and Management Engine holes.
8 of 13 people found this answer helpful. Did you?
No kidding. It's an SC2A11B in the photo of the motherboard.
Does it even make sense to use NVMe? Does the processor have that much throughput?
Only the State obtains its revenue by coercion. - Murray Rothbard
It's not clear what the advantage is over other competing systems
1. Power consumption.
2. Density.
3. Cost
For power consumption, you need to look at both what the server consumes, and the power used to cool the datacenter.
Density means you can pack more into a rack, saving on expensive floor space.
Cost savings aren't clear. TFA doesn't mention prices. But it will likely be less expensive than an Intel system of equal capability.
An editor's job is to get you to read articles surrounded by adverts, not to provide information.
No sig today...
"ARM-based server processors have threatened to take on Intel in the data center for some time ..."
Yes, since about, oh, 2005 was it? ARM servers have been "threatening" and "threatening" and "threatening" for so long, that threatening is all they can do. Or so it seems.
All that threatening must make an ARM CPU tired. Why don't you lay down and take a nap?
Can you really achieve good power consumption (in relation to performance) or density when you're using a six year old processor from 2012? This thing is positively ancient in CPU terms, something like three to generations behind. It's the contemporary of the iPhone 5 or the Galaxy Note 2.
How well is x86 virtualization on ARM these days?
1. qemu-user-mode is a thing and has been for probably the past decade
(And specialized software like DOSBOX has had a great time on ARM gaming devices)
2. Why the hell do you need x86 virtualization on an ARM server for the cloud ?!?
It used to be something important back when most businesses ran Windows servers in-house, running proprietary binary x86 Win32 applications.
Nowadays the cloud is mostly Linux, and is running mostly interpreted languages (like Python, Javascript, etc.)
Linux is opensource, you can compile the user land on any architecture that pleases you (e.g.: look for ArmBian, RaspBian, and other such ARM-specialized Debian derivatives, etc.), and then subsequently Node.js and Python3 will happily execute whatever code you throw at them, no matter if they run on a different Arch than the dev's laptop on which they were written.
If your business use software that you write in an actual compiled-language, you can cross compile or compile on a ARM machine before deployment as part of your devops.
"Sufficiently advanced satire is indistinguishable from reality." - [Tips: 1DrYakQDKCQ6y52z6QbnkxHXAocMZJE61o ]
However, AMD and ARM processors are vulnerable to Spectre.
ARM:
Not all ARM processors do speculative execution. Some cheaper/simpler/lower power/lower cost CPUs do execute things normally (in order, no speculative, etc.)
And that's precisely the case of the Cortex-A53 used TFA's server board. No speculative execution, no way to spectre them.
There's no way to exploit them if the exploited functionality doesn't exist.
(BTW: that used to also be the case on some older Intel Atoms and Xeon Phi)
AMD:
Spectre V1 i.e. "speculative execution working as it is supposed to work" (aka.: the behaviour that was already known to exist and was criticized from day 1 of the invention of speculative execution, but was dismissed back then because "what could you possibly learn by loading things into the cache ?". Cue in cache access timing a few decades later and suddenly it seems a bit more nefarious) is the only one that has been successfully demonstrated on AMD CPU.
And even that one is still an application reading its own memory to which that it already has access to begin with. Nothing freaking and scary, just go learn to keep 3rd party executed code and critically important data separated (e.g.: don't run your password management plug-in in the same web browser process as the JavaScript in tabs that runs any shit you pull from the internet), because you're just one bug away from disaster, be it Spectre or not.
Most of the freaky ones (Meltdown) aren't affecting, because AMD engineers don't have a tendency to throw all security through the window just to scrape a few cycles in some benchmark.
Others (like Spectre v2) haven't been successfully demonstrated in practice: AMD CPU *do* indirect branch prediction, but have much more complex predictors which are difficult to use reliably and any way the CPU speculates a lot less ahead so there isn't that much you can do here, the later being also true for the couple of other variants that also exist on recent AMDs.
TL;DR: there are orders of magnitude of difference between how Intel CPU are affected by Spectre and every body else in the CPU market.
"Sufficiently advanced satire is indistinguishable from reality." - [Tips: 1DrYakQDKCQ6y52z6QbnkxHXAocMZJE61o ]
1. Power consumption.
Except for the weakest home and niche servers, nobody cares about that. It should be 1. Power performance
When all you have is a hammer, every problem starts to look like a thumb.
It may be cool, but what about the 85dB noise level?
I would love the people making these ARM boards to make them for factor compatible with Oracle motherboards so they can upgrade old Sun/Oracle kit. It would be an interesting way to compare performance if you could swap them in the exact location, just switching the cables to the unit above or below.
Sent from my ASR33 using ASCII
It doesn't have any ram on board, the editor is retarded as usual
It has 32GB of eMMC on board and 4 dimm slots for ram
I did a lot of work with "Banana Pi" singe-board computers, and everything thhat comes out with that name, has atrocious hardware support, atrocious hardware and atrocious support.
The hardware is usually barely documented, and does *not* qualify as open hardware, the support forums are censored in Chinese face-saving tradition and full of false promises, and in many, the board design can only be described as *failed*. (E.g. the power regulator not supporting the storage power requirements when powered via USB port. Or the storage and gigabit network interfaces maxing out at 400 MB/s even with tuning.)
Unless you want to be in a world of madness and pain, don't buy them.
(And I'm not one of those who think that China is crap in general. I was very happy with my Blackview BV6000, in all aspects.)
mate on a server? and weak IO as well.
only 2 sata and pci-e X16 at X4??
Not really. Data centers spend a lot of money cooling their systems. If they can get a denser, more power-efficient setup that can do the same thing, they'll definitely be interested in that.
The demo shows building a Linux kernel but how long does that take? How long does Yocto take to build on such a machine?
Its the fun of moving existing quality code from "Intel".
Showboating on a new type of system rather than just been another person working on a large exiting projects.
Who wants to be part of a very large project when new projects await on ARM?
Domestic spying is now "Benign Information Gathering"
You have a certain amount of work to do. You want to get it done using the least amount of power possible. That is power performance. That is also how you spend the least money cooling systems. See now?
When all you have is a hammer, every problem starts to look like a thumb.
Thomas Dorr made an interesting note on this project.
According to Thomas's post this looks to be similar to the following:
https://www.96boards.org/produ...
With a few modifications this could look to be ARM SOC's being put into mainstream PC use.
Then I misunderstood what you meant. "power to performance ratio" or something like that would have been clearer. The phrase "power performance" just sounds like you're talking about souped up gaming machines or something where you don't care about power usage, just performance.
Let's call it energy efficiency, the standard terminology.
When all you have is a hammer, every problem starts to look like a thumb.