Ampere eMAG 64bit Arm Workstation Enables Native Arm Development

Over the last few years, several companies have come up with 64-bit Arm workstation to allow developers to test Arm code natively which may be important to avoid network delays or test applications requiring video or graphics.

Those started to become available in 2018 from the relatively low-end  24-core Cortex-A53 Linaro “Synquacer” Developerbox to the much more powerful (and expensive) GIGABYTE ThunderXStation Workstation powered by up to two 32-core ThunderX2 processors.

In the fall of 2019, SolidRun started to ship HoneyComb LX2K 16-core Arm Workstation motherboard with and NXP LX2160A 16-core Cortex-A72 processor that offers significantly more performance than the Linaro Box at a reasonable price ($750). While reading a recent Anandtech post with photos of an engineering sample, I found out Avantek was also offering the Ampere eMAG 64bit Arm Workstation powered by an Ampere eMAG 8180 32-core server processor.

Ampere eMAG 64bit Arm Workstation
Source: Anandtech – Click to Enlarge

Ampere eMAG 64-bit Arm workstation specifications:

  • SoC – Ampere eMAG 8180 32-core Armv8 processor @ 2.8GHz / 3.3 GHz (Turbo) with 3MB L3 cache; 125W TDP; TSMC 16nm FinFET+
  • System Memory – 16x DIMMS 2400/2666 MHz, up to 512GB.
  • Storage – 3x SATA3 HDD/SSD, 1x M.2 NVMe PCIe x4 SSD
  • Video Output
    • VGA built-in
    • Optional discrete graphics cards
      • AMD FirePro W2100 2GB 2x DP PCIe x16 gen 3
      • AMD Radeon Pro WX 5100 8GB DDR5, PCI-E x16, four DisplayPort, 1792 Stream Processors, 3.9 TFLOPS
      • Nvidia Quadro GV100 32GB
  • Networking
    • Gigabit Ethernet (RJ45) via Intel i210 controller
    • 1x 10/100M management LAN
    • Optional add-in-cards (no details provided)
  • USB – 4x USB 2.0 ports
  • Serial – 1x Serial RS232
  • Expansion – 1 x PCIe Gen3 x16 Slot (8 lanes)
  • Misc – AMI AptioV BIOS, Aspeed AST2500 BMC
  • Power Supply – Corsair VS650 650Watts power supply
  • Dimensions – 577 x 243 x 585 mm
Ampere eMag Arm Developer Box
Source: Anandtech – Click to Enlarge

Supported operating systems including Ubuntu, Centos, SUSE SLES Arm developer edition, and openSUSE.

Price starts at $3,391 with the minimum configuration including 32GB DDR4 ECC RAM (4x 8GB), and a 240GB SATA SSD. But from there it also depends on your requirements and budgets, with a maxed-out configuration going for around $25,000.

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

47 Replies to “Ampere eMAG 64bit Arm Workstation Enables Native Arm Development”

  1. > Price starts at $3,391

    …excluding a graphics card. No display output could be a bit problematic for this unique type of device (data center SoC on a stripped down server board inside a PC tower).

        1. Great, as expected no possibility to extend the multi 1000$ “workstation” soon even the H6 will have more flexibility…

          The nxp based board was much more flexible with the plenty of configurable serdes lanes, not too sure how it would hold up as a workstation.

      1. This is the mirrored output of the AST2500 BMC. This BMC is nice to remotely administrate your ‘workstation’ (keep in mind that this is a server board fitted somehow in a PC tower enclosure).

        But I fear if you think in 2020 a single analogue VGA output would be suitable for any serious and especially developer work (requiring multiple high resolution displays these days that do not flicker like hell) then…

        1. Many (most?) higher-end workstations come with no on-board graphics and Intel higher-end Xeon CPU don’t have graphics IP.

          1. I know. But why are we talking about ‘workstations’ since this is just a server board fitted into a PC tower?

            Server motherboards usually have a BMC (Baseboard Management Controller) that almost always also acts as a primitive 2D GPU and allows to access BIOS/UEFI remotely (KVM over IP). The output is also mirrored to a VGA port since most KVM switches still use VGA but it’s completely useless for anything you would do on a workstation.

          2. Lets just use things for what they are intended. More than os installation and bios/efi stuff noone would do on a real server (not talking about soho windows servers)

          3. I’d do boinc, data processing, Java, web browser stuff, html 5, internet server stuff (like photo/video server)…

          4. I’ve already coined the idea to AMD, and was rejected/ignored. In the mean time, Intel took up on the offer. What they need with so many cores, is larger contact area with the CPU lid, for better cooling. Having a mini GPU is a perfect solution! Meaning, something that can display a 1080p gui or terminal over DP or HDMI. Not something you’ll want to watch movies with, or even game on. While the GPU is small, it also increases surface area to cool the CPU, and in the server market is mostly turned off when not used, meaning, it won’t add any heat to the CPU.

            If you want 20 screens, buy the pcie hardware that supports it!

            Another option was mobo manufacturers building a weak ‘igp’ on the Mobo itself, utilizing 1 to 4 pcie lanes.
            Intel is working on implementing a tiny igp in their higher core xeons!

  2. Impressive I/O bandwidthh in this SoC, which would better express its power in a server than a workstation, though it’s probably quite beefy there at 3+GHz. The TDP at 4W/core looks quite high for ARM chips but reasonable compared to similar-class Xeons. As usual the price will keep it into niche markets.

  3. And a budget range $75 non modular PSU… What a joke.
    They couldn’t even do a decent job with the cable management at this price.

  4. Anybody saw working nvidia/amd GFX card with ARM based SBC ?
    Some SBC has PCIE but I didn’t saw any with working big GFX card 🙂

    1. How many SBC do you know that come with a PSU and power circuitry to cope with +75W for something the ARM SoCs on these SBC already have integrated in a far more efficient way?

  5. Too expensive ! in the world that AMD 3970x exist and can be found in similarly priced system. (And other EPYC workstations)

  6. ARM will remain at a massive disadvantage as long as no companies are selling reasonably priced developer boxes with good single thread performance. More cores generally equates to lower single thread performance. So this product is a failure in the general sense because it’s an attempt to make a competitive ARM HEDT while completely disregarding what developers actually want in a workstation.

    I’d still take this over any prior “ARM Workstation” because this has better single-thread performance than anything else I’ve seen in that marketplace… if you can call it a marketplace yet. But no corporate is going to buy this for their developers, and I sure as heck wouldn’t pony up my own dollars for it… so while it’s the best available, it’s still not going to see any sales volume.

    When they start to sell “workstations” at around $2000 with 4-8 cores that outperform an iPhone, then they will actually get taken seriously.

    If ARM wants to succeed, they should be spearheading designs the same way Intel has when they have wanted to enter new markets.

    1. Sorry if my question might be stupid, but i’m not the targeted audience of such a board.
      But why would single thread performance make a difference? I thought compilers use as many cores as they get if they are invoked properly?

      I always thought single thread only matters with unoptimised home user sw?

      1. Just because most operations you perform on a developers machine are limited by the time to complete a task. You run “./configure”, it’s full of single-threaded operations. A browser takes ages to render a page, idem. You search for some contents in your mailbox, ditto.

      2. I think you are confusing a developer workstation with a build server

        Whenever there is user interaction, good single thread performance is important. If you look at the x86 world and how it splits its CPU designs up you will see that Workstation class CPU’s (AMD TR, Intel Xeon W / Enthusiast) have much higher single thread performance than the server class CPU’s (AMD Epyc, Xeon Gold / Platinum). Usually workstation class CPU’s match, within a few %, the ST performance of the highest end gaming CPU’s from the same stable

      3. Because modern developers are working with code bases in the tens of millions of lines of code, and our IDEs need to be able to perform indexes, searches, static analysis, in real time on code bases that large. And for the most part that is still a single-threaded job.

        When you’re doing iterative development (popular these days) your round-trip-time for testing cycles to prove out your modifications is a significant throttle on your productivity. If it takes 2 minutes to do a build and you spent an average of 2 minutes between compiles, there’s 50% of your productivity out the window. Multiple cores can help with that, not quite as well as faster cores. I agree it’s better than nothing. But if you add an extra ten seconds waiting for code completion of class methods and exception handling, and various other lookups… heck if Google Chrome runs 30% slower on your workstation and you’re just visiting StackExchange a lot… that will hurt your productivity, and try your patience, especially when your work environment is slower than the phone in your pocket.

    1. Of course it is.

      First step: choose an ARM CPU for the task. You need high single-threaded performance, somewhat decent IO capabilities (PCIe/USB3) and a reasonable amount of fast memory. Which ARM CPU will you take? Wrt performance I can only think of Apple’s A12/A13 but PCIe situation and amount of addressable RAM most probably sucks. And of course you won’t be able to buy these SoCs and even if Apple sometimes in the future might switch to ARM with macOS a ‘PC style device’ will be the last thing considered.

      So there’s no mobile or desktop CPU but only expensive data center and server ARM CPUs left. Developing a motherboard for a non existent market is a challenging task since ‘economy of scale’ will not work here. So you end up putting your overpriced eMAG 8180 server CPU on either a (stripped down) server board as it happens here or you do the Solid Run approach based on a LX2160A SOM (developed for platforms for which markets exist).

      1. Well, the SBCs themselves (including the Raspberry Pi) once also were something “for a non-existent market”.

        And it doesn’t need to be “high” performance, there are x86 boards like
        Can’t be too hard to put in a 4-8 core arm64 there for the SoC.
        Of course the pipe dream would be the above, but with RISC-V instead.
        But let’s at least get something at all with RISC-V, first.

        1. > Can’t be too hard to put in a 4-8 core arm64 there for the SoC

          With which use case in mind? If it’s a developer workstation then the SoC needs to have high single-threaded performance. And also a decent amount of RAM and some IO capabilities are needed. So which SoC to choose?

  7. Now, that’s some festival of greed, if you don’t mind me saying.
    For that price it looks like an advertisement for cross-compiler.

    1. > that’s some festival of greed

      Huh? Do you have an idea how much Ampere is charging for the eMAG 8180? Did you realize that this thing is made for an absolute niche requiring to mechanically fiddle around with all components to end up with a boring PC tower in the end?

      Quoting the aforementioned Anandtech post: ‘Avantek, who takes one of the Ampere eMag motherboards and places into a consumer based PC chassis from Be Quiet, then modifies the chassis to fit the server-sized motherboard. This involves customization, given that the server motherboard does not have the standard E-ATX holes or PCIe spacings for the rear panel’

  8. They really should stick to selling motherboard and processors. It makes no sense to pay more than $1k for a Mobo/CPU combination (and that’s for the 32 core CPU), and leave the accessories, like case, PSU, ram, SSD, and GPU to the end user.
    With that closed mentality, and 3x to 25x their market price, it’ll really be a hard sell!

    Companies want to save money. It’s competitive with an Epyc system at 3k, but not competitive with a Ryzen system.
    Companies would rather want to buy a ryzen with non registered DDR, than Epyc or this at 3 to 20x higher price, if it means their servers need occasional rebooting. Mine runs fine for a good month 24/7, before a reboot is needed. Sometimes even longer.

    1. While you wouldn’t use an ARM server for quite the same things (e.g. no developing drivers on them) I was also curious. The biggest AWS EC2 A1 instance has 16 cores and 32GB of RAM so let’s get two of those and bump the eMAG up to 64GB to keep cores and RAM matched overall. The eMAG now costs $3938.40 and the instances cost $0.9216 per hour. That’s 4273 hours of usage paid for, or 534 days of 8 hours a day usage, or 178 days of 24/7 usage.

      I’d probably buy an LX2K instead of either option.

      1. > The biggest AWS EC2 A1 instance has 16 cores and 32GB of RAM so let’s get two of those

        Why in this context (locally working on something not dog slow but with a specific ISA)? And why choosing an A1 instance with the significantly slower old Graviton at 2.3Ghz instead of M6g? Anyway, if you really want to compare I would choose the very same ARM CPU and that’s then a c2.large.arm instance:

        1. My bad, I didn’t realise they’d changed the instance naming scheme for graviton 2. A2 should have followed A1. That said, looks like m6g isn’t generally available or featured on the instance pricing overview yet so let’s do as you suggest and switch to a directly compatible server.

          $1/hr from vs $5,156.64 for an equivalently specced eMAG, that gives you 644 days usage at 8 hours a day or 215 days at 24 hours a day.

          1. > $1/hr from vs $5,156.64 for an equivalently specced eMAG


            If you do some server stuff somewhere in the cloud that’s an entirely different use case compared to what the featured PC tower here is about (working locally on an ARM ‘workstation’ for whatever obscure reasons).

            If it’s about server tasks I would compare with a server, e.g. Lenovo ThinkSystem HR330A (and then I would also factor in all the other costs associated with running your own server infrastructure).

  9. This is all about the arm environment.
    Arm is a CPU, yet nowhere do I see it trying to be competitive with x86 in terms of cost and performance!
    How easy is it, to just burn more arm cores on one package, and end up with a 65 or 95 watt CPU?
    I see all kinds of devices, from routers to print servers, etc.. but not a single compute unit!
    There’s no reason an arm CPU with 32 cores should cost more than $240!
    I mean, for 240, you’ll get 8 AMLogic TV boxes, that not only come with quadcore cpu, But also with Android is, remote, power adapter, lan and wifi module, audio, 4GB of emmc and 1GB of ddr3 ram PER UNIT!
    $<250 is more than a fair price for a 32 core arm cpu, and $200 is extremely generous for a motherboard!
    Once these things get sold in bulk, prices could go as low as $120 for a 64 core CPU, and $75 for a motherboard (that doesn't have to have more than budget x86 mobos from gigabyte or Asus or ASRock).

    1. I suggest because theiir is no bulk large scale market yet for what you describe. ie milions units sales.

      I will get down voted, but history shows its consumer use that drive advancement and costs down.

      The Nintedo Switch helped Nvidia afford the Nvidia Shield TV SoC upgrade.

      Now Amlogic, Rockchip, Nvidia and maybe re-used Mobile SoC for Linux laptops, desktops, handhelds or Chromebooks, desktops, kiosks and terminals would help..

    2. There was once up on a time…
      When you could (at least in theory) buy a simple board for an athlon and plug a dec alpha in that very board.

  10. A ton of comments bc this is a dinosaur. I bet the dev team was a bunch of boomers. Sad, but some truth to this.

    1. Wow, thats some delusion for you. Thankfully we are usually free of this kind comment here but thanks for showing us the calibre of your brainwashed mind Jay

Leave a Reply

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

Khadas VIM4 SBC
Khadas VIM4 SBC