Cavium ThunderX based Scaleway ARMv8 Cloud Servers Go for 2.99 Euros per Month and Up

Scaleway launched 32-bit ARM server hosting services in 2015 for 10 Euros per month, before dropping the price to 2.99 Euros per month half-year later, and now the company has just launched a new offering with 64-bit ARM servers powered by Cavium ThunderX processor going for 2.99 to 11.99 Euros per month depending on configuration.

Click to Enlarge

The processors are equipped with DDR4 ECC memory, and all three services included unlimited transfer, so you don’t need to pay for any bandwidth fee. While the price is shown per month, you’ll be billed by the hour (0.006 Euro/h for ARM64-2GB), so if you are using those for development it may even cost less per month, as you can turn them off when not working.

All server are located in a Paris data center in France, and runs Ubuntu 16.04, but more operating systems and “InstantApps” will be added to the selection. More servers will soon be available in their Netherlands datacenters (AMS1).

You can add a new ARMv8 server in Scaleway dashboard to get started with the new servers. You’ll find a few more details on Scaleway Virtual Cloud Servers page.

Share this:
FacebookTwitterHacker NewsSlashdotRedditLinkedInPinterestFlipboardMeWeLineEmailShare

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

ROCK 5 ITX RK3588 mini-ITX motherboard

22 Replies to “Cavium ThunderX based Scaleway ARMv8 Cloud Servers Go for 2.99 Euros per Month and Up”

  1. I was using their C1 baremetal servers back in the day, and they were fine, but I moved on to ARMv8. I’m glad they’ve upgraded to ThunderX – they have a returning customer.

  2. The price for a similar x86_64 setup is the same 2.99 euros.
    I wonder what is the best choice, performance wise??

  3. @JotaMG
    You can pretty much guarantee it’ll be the x86 servers that win on performance. ARM has advantages, cost and efficiency, so it should be cheaper to buy and run for Scaleway. You’re an end user though, you shouldn’t care about power consumption or much of anything except performance and price. If lower operating costs for Scaleway don’t translate into lower prices for you there’s no benefit to ARM. What we’d need is someone like Amazon picking this up since they’ve got the volume and a long-term view.

  4. @fkpwolf
    I think baremetal here means you had access to your own hardware with the C1 instances.
    With the ARMv8 Cloud Servers, you share the hardware with others, but you still have your reserverd 2 to 8 cores. Cavium ThunderX (v1) processors have 8 to 48 cores each.

  5. @fkpwolf
    It isn’t bare metal. The C1 instances are bare metal (32 bit). The Cavium ones are virtual servers because the Cavium could have 48 cores. There are also Cavium CPU with 8 Cores available but I don’t think they using it for the server. The density of core per rack unit is not high enough.

  6. @CampGareth
    The x86-64 server are based on Intel Atom Avoton SoCs (not sure which exact yet), which are pretty slow for x86 processors.
    Maybe it would deserve some comparison against ThunderX, but again we don’t know the exact. People who tried the servers may give some more details here.

  7. @JotaMG
    Don’t jump to conclusions too quickly. Some x86 “cores” are actually threaded cores, not physical cores. Some providers also further split cores to make them all perform the same on machines with different levels of CPU performance. I’m not sure what these people do but the only way to really know is to benchmark. You also have to benchmark the things that matter for the software you’re running, since each system will have differing memory, disk and network performance, each one of which may be the determining factor for your setup.

  8. Re performance of the new servers, here’s something posted on the Scaleway forums by user Bob:

    I did some sysbench CPU test runs (10000 prime numbers). It’s a very basic test, but here are some ballpark figures:

    x86-64 “Starter” (VC1), Atom Avoton C2750, 10000 primes in 9s on one core
    ARMv7 “BareMetal” (C1), 10000 primes in 258.5s on one core
    ARMv8 “Starter” (ARM64), 10000 primes in 6s on one core
    x86-64 “Intensive”, not tested

    This is a very parallelisable task, so e.g. 2 cores took half the time of 1 core. I suspect one of the operations used was particularly bad for the ARMv7 core, since the result is very poor.

    Of course, you must consider your specific use case – this is a very simple benchmark. You can test yourself at the low €/hour rate Scaleway charges…

  9. @blu
    When do people stop using such moronic anti benchmarks like ‘sysbench’? This thing runs in L1/L2 cache only (great if your application does not and is affected by memory throughput) and calculating prime numbers is also nothing that relates to any real workload!

    This sysbench anti benchmark finishes stuff magnitudes faster when running on some platforms (that’s why ARMv7 looks that bad above) since it’s not a hardware test but one for compiler optimizations. And when you use sysbench on exactly the same hardware numbers with Ubuntu Xenial are 30% better than those made with Debian Jessie when using distro packages since… GCC 4.9 vs. 5.4 (yes, that’s all — really).

    And if you want to lower your scores just use sysbench 0.5 instead of the widely used 0.41 everywhere: performance now drops by 10 percent. So if you upgrade the benchmark it reports your hardware got slower!

    TL;DR: sysbench is crap for various reasons to measure CPU performance. It’s great to learn about compiler optimizations though. And never trust in any sysbench numbers on the net unless sysbench version and GCC version is known! Otherwise numbers are not comparable.

  10. – “openssl speed rsa2048 -multi ”
    – 7z b
    – minerd –benchmark

    and you’re done to get a rough estimate about CPU horsepower in short time (focussing on important stuff like integer, float, memory performance so you already have an idea how/whether the numbers spitten out by these tools will correlate with your application!). Using distro packages you know then what get if you’re lazy. Firing up the compiler with appropriate settings and you might end up with numbers twice as good (for free, just by letting the compiler do its job optimizing for the platform in question)

    But on such servers IO bandwidth/latency is also important so another round with fio (try out several mixed read/write loads) and ioping is necessary.

    Using sysbench’s cpu test is only fooling yourself (unfortunatly that’s done everywhere, same with dd and hdparm for ‘storage benchmarks’)

  11. @tkaiser
    I’m of the opinion that no synthetic benchmark is definitive, and also that no synthetic benchmark is absolutely useless – they all demonstrate *some* aspect of the performance of the system, where the latter often includes the compiler, as that is often insurmountable – real sw does get largely affected by the quality of the compiler and its settings. I’m far from the thought that a prime-number generator is indicative of the overall performance of the system, or, say, a particular task other than the one being measured ; )

  12. I did a very basic speed comparison with a notebook (an apple macbook pro with an old I5) running my java code. ARMv7 “BareMetal” (C1) has around 70% of the performance of my macbook. HOWEVER this is not a real speed test. Perhaps the JVM on ARM is more efficient, perhaps their SSD are better than my SSD. Perhaps Ubuntu 16 is faster than Sierra, etc, etc. Both machines are too different and my “test” is too cheap to use it as a serious comparison.

    3 euros for 70% of the speed of my little apple is good enough for me and my server though. I suggest you test it. You can buy some hours and run some tests.

Leave a Reply

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

Khadas VIM4 SBC
Khadas VIM4 SBC