28

I was interested in the specs of the UNIX server my university provides for students, so I ran screenfetch. Here's the output:

                                               user@unix4.university.edu
              `.-..........`               OS: Red Hat Enterprise Linux 7.5 Maipo
             `////////::.`-/.              Kernel: x86_64 Linux 3.10.0-862.14.4.el7.x86_64
             -: ....-////////.             Uptime: 9h 1m
             //:-::///////////`            Packages: 3796
      `--::: `-://////////////:            Shell: bash 4.2.46
      //////-    ``.-:///////// .`         CPU: Intel Xeon E5-2680 v2 @ 40x 3.6GHz [61.0°C]
      `://////:-.`    :///////::///:`      GPU: Matrox Electronics Systems Ltd. G200eR2
        .-/////////:---/////////////:      RAM: 8290MiB / 64215MiB
           .-://////////////////////.
          yMN+`.-::///////////////-`
       .-`:NMMNMs`  `..-------..`
        MN+/mMMMMMhoooyysshsss
 MMM    MMMMMMMMMMMMMMyyddMMM+
  MMMM   MMMMMMMMMMMMMNdyNMMh`     hyhMMM
   MMMMMMMMMMMMMMMMyoNNNMMM+.   MMMMMMMM
    MMNMMMNNMMMMMNM+ mhsMNyyyyMNMMMMsMM

All I can find about Matrox GPUs is their wikipedia page which says that the G200 was released in 1998. Why would my university have them in a modern server (CPU was released in late 2013)?

  • 9
    Do you think a UNIX server should have what? A Nvidia card? For what games exactly? No, a server just needs something to display text mode most of the times. I don't remember but people back in 1998 were already using graphical desktops, with Windows 98? A Matrox G200 is way more powerful than needed for a server. – GabrielaGarcia Nov 2 '18 at 20:05
  • 12
    @GabrielaGarcia A lot of students use this server for CS homework and I'm in a class using Tensorflow. I was hoping there would be some CUDA GPU available to play around with. – user960076 Nov 2 '18 at 20:07
  • 4
    I am sure there are CUDA special processors, just not on that server. They cost a LOT of money (and no, what you likely know as CUDA is not even funny - NVidia does special cards, wait until you get access to a server with 6 of those for Tensorflow work... each eating a 2080ti for breakfast). – TomTom Nov 3 '18 at 11:35
  • 4
    @trognanders Hypervisors don't emulate a Matrox (and nobody would virtualize one either, if at all possible). They either have their own GPU driver or they expose a generic VESA SVGA card. – user71659 Nov 4 '18 at 2:52
  • 3
    @trognanders: I have several physical servers from Dell and HP here, and most of them use G200, attached under the BMC's own PCI bridge as user1908704 has mentioned. (Although the oldest one uses ATi ES1000.) – grawity Nov 4 '18 at 9:59
71

General-purpose servers don't need a modern GPU - just enough to show a medium-sized console desktop. They mostly deal with regular CPU computing and networking.

Matrox G200 VGAs, however, are commonly used on servers due to their integration with a baseboard management controller (BMC, also known as iLO, iDRAC, or the IPMI).

This management controller acts as an independent system with its own operating system and lets the server's administrator remotely connect to the console display & keyboard – they can see the BIOS screens, restart a server even if it's completely frozen, even start it from full power-off. For these tasks, the controller must know what the graphics adapter is displaying right now.

So I would guess that the old Matrox video adapters are used for this because they store the video buffer in system RAM (instead of their own VRAM) and use a sufficiently simple data layout that the BMC can decipher it without needing arcane knowledge about the GPU's internals, nor without any help from the main OS.

(Or perhaps the opposite – as mentioned in comments the G200 is usually built into the BMC, possibly giving the BMC completely direct access to the G200's video buffer.)

But even if the server was built for GPU computing, I assume it wouldn't have an "all-in-one graphics card" as PCs do, but instead a set of dedicated compute-only GPGPUs (e.g. from nVidia) for the heavy work – and still the same Matrox VGA for the console.

  • 23
    I suspect it's more likely that, when Matrox became irrelevant to the video card market, they sold or licensed the G200 design for whoever produced the BMC to include in the BMC chip, integrated with the remote administration functionality. – R.. Nov 3 '18 at 1:00
  • 14
    There's no G200 chip, it's implemented inside the BMC, which already speaks PCIe. I had assumed it was actually a reimplementation of the register-level interfaces (in the way that all GPUs can do basic VGA without being based on something from 1985), but it turns out Matrox do licence the G200 IP to chip designers. – user1908704 Nov 3 '18 at 14:34
  • 4
    In addition the drivers for the G200 and similar Matrox "chips" are old and stable and rock-solid. That too is important on a server. – Jamie Hanrahan Nov 4 '18 at 3:10
  • 8
    @R.. - "when Matrox became irrelevant to the video card market" ... Matrox have never become irrelevant to the video card market. They've simply become specialised -- AIUI, they are the market leaders in producing video cards for systems that need to support very large numbers of monitors. – Jules Nov 4 '18 at 7:04
  • 2
    @JamieHanrahan: If only the whole BMC also were rock-solid, too... – grawity Nov 4 '18 at 9:36
34

That Matrox G200eR2 is not a separate video card. It is a chip directly integrated into the server motherboard. It is cheap, very reliable, easy to integrate and provides excellent text (console) display ability and decent 2D graphics ability. Is is also so well known that just about every Operating System for Intel hardware has driver support build-in for it.

The only purpose for a VGA card there is to get a basic console display that you can use for Bios setup and initial installation of the server. After that you will probably only ever access the server remotely. It doesn't have to be a good VGA card. You are not going to be gaming on it. But it is a major blessing if it works out-of-the-box with whatever OS you are going to install on the server. And that is all you need and want in a server.

Matrox chips have always been very popular for this purpose and this particular one was still used in 2014 in new Dell servers and probably in some other brands as well.

  • Actually, on most servers the (graphics and text) display support is there only for operating systems like Windows. The management interface almost invariably supports BIOS setup via serial port (usually accessed through the BMC via the network) and OSes such as Linux can do a full install and subsequent operation with a serial console, meaning that you never have to use the "video" console display at all. This allows remote installation of a bare machine you've never even touched, though the same can often be done via remote video over network. – Curt J. Sampson Nov 3 '18 at 2:39
  • 2
    You are not going to be gaming on it. - speak for yourself, i made automation programs for online mmorpg games and had it running 24/7 doing rudimentary mundane but beneficial tasks, on servers. – hanshenrik Nov 3 '18 at 8:59
  • @CurtJ.Sampson, I'm not sure about that... A VGA output is still the "default" interface, which every device has. With it, you can connect completely different devices to a single KVM switch (network-based or not), and for the really simple cases you don't need to touch the management controller at all. Regardless of the OS, not having a VGA output and a keyboard connector would be a bit of a nuisance. – ilkkachu Nov 3 '18 at 10:05
  • 1
    @ilkkachu VGA is the "default" for "PC" ISA hardware, but never was for most other server systems (Sun, etc.). Anyway, if you find it convenient to use a KVM or whatever, sure go ahead with that, but straight serial has a lot of advantages (e.g., faster, esp. over bad networks, you can log all the output, more easily automated, ...). As for the management unit, do not ignore it unless you don't plug in its network port. Even when well-configured they are incredibly insecure, and an unconfigured one invites a system takeover by anybody who can send it packets. – Curt J. Sampson Nov 4 '18 at 4:05
  • Oh, and I can say from extensive experience that, for Unix servers in typical professional environments, not having VGA is not in any way a nuiscance. In fact, having it, and having to ensure the system doesn't use it, can be more of a nuisance. (You don't think all those cloud servers that Amazon and Google run use real or emulated video output, do you?) – Curt J. Sampson Nov 4 '18 at 4:07
15

Why would my university have them in a modern server (CPU was released in late 2013)?

Because a server does not need a high-performance GPU.
And by the way, Matrox had good Multi-Monitor graphics cards long before ATI/AMD and NVidia had them.

So the decision had probably been logical by the time of purchase.

  • Very logical and very cost effective, I assume, given the chip's age. – GabrielaGarcia Nov 2 '18 at 20:09
  • +1 for old trusty Matrox'es... – einpoklum Nov 4 '18 at 23:48

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy