8 votes

I would also like some PC build advice (workstation build)

Hi all. I'm working on an idea for my next desktop/workstation build since I stumbled into some resources that make something more awesome than what I currently have a reasonable option.

What I want to use (because I already have these and thus can save money - got a bunch of rackmount computers at auction and got pretty lucky): infinite sticks of 16GB DDR4 3200 (ok, so I have 20 sticks left after selling a bunch, not exactly infinite but good enough for wanting a single maxed-out build) and 4 RTX A4000 16GB cards. Probably I only intend to use 2 of those and sell the other 2. I also am spoiled for options with various 4U rackmount server cases and a couple of them have 1,000W power supplies (the 2 computers that each had 2x A4000 cards had 1kW power supplies).

My current desktop is an i9-11900k / 64GB DDR4 3200 / RTX A2000 6GB. At first I was interested in just popping a couple of the A4000 cards in but then research banged my head into the issue of PCIe lanes being limited by the CPU.

Options I'm currently considering are Threadripper Xen 3 generation (since that uses DDR4 3200, lets me use 8 sticks depending on motherboard, and gives me 128 PCIe 4.0 lanes), EPYC of some variety since some of those also fit into the 'can use tons of RAM and have tons of memory channels + more than enough PCIe lanes' or..... dual Xeon? AMD seems like the better choice though - but correct me if I'm missing something there.

If I've missed any other options that would give me enough PCIe lanes for at least 2 A4000 cards both connected at full PCIe 4.0x16 and the ability to utilize as much of the RAM as I have (and have the advantage of more memory channels than standard desktop processors) please let me know.

Also looking to not break the bank too much, something that fits into the sub-$2K range, or close to that, would be preferable (and cheaper being more preferable). Anything that goes past $3.5K is almost certainly not up for consideration.

Last point of consideration is that I'm not looking for any part of the new build to be less capable than what I already have, and am not looking to spend significant money for a non-significant capability increase.

Use-cases relevant to my desire for the horsepower: video rendering with Blender (I certainly wouldn't mind rendering going a lot faster when outputting H.265 or rendering hour+ videos), and dipping my toes into local AI (have been interested in this but until I lucked into the A4000 cards it wasn't really possible. With a couple A4000 cards I should be able to run some of the lighter-weight models, at least I believe so (I have, as of yet, no experience running local AI). Would be running Linux (I currently use Linux Mint). Gaming is an afterthought and I don't play anything AAA / ultra-high graphics. Mostly for gaming I just enjoy playing Magic: The Gathering Arena as I drink my morning coffee, so no major hardware requirements there.

Thoughts?

Edit: the DDR4 3200 is non-ECC RAM.

5 comments

  1. [2]
    goose
    Link
    Beware, as I learned this the hard way: ECC RAM is not compatible with most consumer motherboards. They may have the same DDR spec and physically fit in the slot, but you'll fail to POST. If your...

    got a bunch of rackmount computers at auction and got pretty lucky

    Beware, as I learned this the hard way: ECC RAM is not compatible with most consumer motherboards. They may have the same DDR spec and physically fit in the slot, but you'll fail to POST.

    If your server RAM isn't ECC, or if the motherboard you're plugging it into can accept ECC, then you're good.

    2 votes
    1. l_one
      Link Parent
      The RAM is all non-ECC. Standard DDR4-3200.

      The RAM is all non-ECC. Standard DDR4-3200.

      2 votes
  2. [3]
    Tukajo
    Link
    I know you're not looking for this but, I've got almost an identical situation to you. 64gb DDR4, 3090 ti. I use Linux Mint. I have and do mess around with local AI. If you want some advice, reach...

    I know you're not looking for this but,

    I've got almost an identical situation to you.

    64gb DDR4, 3090 ti.

    I use Linux Mint.

    I have and do mess around with local AI. If you want some advice, reach out to me. I can at least provide you with some setup and workflows

    I've done some pretty useful stuff with Local AI from homelab automation, to actual programming. I'm a software engineer by day.

    Would love to talk shop even if it's just "hey what's do you wish you'd have here?"

    Or "what model/setup should I use here?".

    1 vote
    1. [2]
      Viceroy
      Link Parent
      Your setup is super similar to mine, I have 64gb of DDR5 but only a 4070 super. I run fedora and have been interested in playing around with local models for my homelab but just haven't figured...

      Your setup is super similar to mine, I have 64gb of DDR5 but only a 4070 super. I run fedora and have been interested in playing around with local models for my homelab but just haven't figured out what to do with them. What workflows have you introduced or what has been the most fun to play around with?

      1 vote
      1. Tukajo
        Link Parent
        I've had good luck with handling sensitive data (think structured data that is highly private but annoying to locate, or maintain/deduplicate). I've also had luck with automating some research on...

        I've had good luck with handling sensitive data (think structured data that is highly private but annoying to locate, or maintain/deduplicate).

        I've also had luck with automating some research on my proxmox clusters. Things like, identifying structures that could be "better utilized" (system resources used in a way that is maybe inefficient).

        I also had a lot of luck driving it to debug, reverse engineer, and speed up Civilization 4 and a mod I play, because it was unbearable and broken on modern systems.

        I use some custom skills to do this, mostly with "OpenCode" as my agent harness. I've heard others like Pi and Hermes. I haven't used the others.

        If you are looking to host models, I HIGHLY advise a setup with the "Strata" Engine.

        It's a new approach that uses Qwen 3.8 Next, to have the FULL model running on systems like ours, where "high frequency" layers are on VRAM, and low hit layers are offloaded to system ram.

        For me, the heaviest version uses my entire 24gb vram and about 60/64GB of my system ram.

        As someone who puts up with AI daily at my regular job, this model is nearly as good as frontier Opus 4.6 was 6 months ago. Which is kind of insane. It does everything I ask of if, albeit I'm more technical with my descriptions, which I'm sure helps.

        Happy to provide more detail if you'd like!

        1 vote