Building an AMD-based PC

34 thoughts on “Building an AMD-based PC”

1. Blog related topic, #divorce: Kentucky divorce rates dropped 25% after 50:50 custody became the norm.

   https://lavocedinewyork.com/en/news/2025/09/07/kentuckys-shared-custody-law-revolutionizes-divorces-but-raises-alarms/

   Reply

   • Thanks, Anon. This confirms the classic “‘These boots are made for walking’: why most divorce filers are women” paper (Brinig and Allen 2000) https://academic.oup.com/aler/article-abstract/2/1/126/141123

     PDF at https://www.bereanpatriot.com/wp-content/uploads/2019/04/Divorce-Reasons-Study-These-Boots-Are-Made-For-Walking.pdf

     (Women sue because they expect to win custody of the kids and all of the cash and noncash benefits that go with that victory.)

   • I came here to post this article too. Would love to see a dedicated blog post about it.

     The WSJ coverage of the topic seemed to focus on on edge cases where there is actual physical abuse by the husband against the wife and didn’t cover at all the chance that the abuse might be made up (i have experienced this first hand).

     If we had this default 50/50 custody deal in my home state i could have saved myself $100k+. In the end i had to appeal my custody case to a higher court (screw JDR court) and ended up winning but it took years and a pile of cash that i will never get back.

     As far as PCs go why not go with a linux distro on your current PC and kick windows to the curb?

2. Re: built-in UPS

   Comes standard with every laptop and smart phone. I never bought one because they used lead-acid batteries. Maybe they are li-ion now? We were thinking about a natural gas backup generator, our power goes off for days (water too after an inland tropical storm wiped out our supply). I bought a new laptop with Windows 11about 3 months ago for $350 on sale.

   Re: DEI and intel

   Intel mentioned rolling back some of it’s DEI “progress” in its last annual report, joining the DEI-reversing bandwagon. Also, the U.S. just invested $9B for a 10% stake in Intel with some small change it found in the couch (unused Biden-era “CHIPS and Science” funds).

   “If you can’t enlist–invest–buy a Liberty Bond [or Intel processor]. Defend your country with your dollars.” — WWI U.S. War Bond Poster

   Reply

3. Still a lot of money for a blog commenter. For that kind of money, who is actually going to use it? Greenspun doesn’t live permanently in FL but is a digital nomad.

   Reply

4. HDDs redundantly configured in RAID for reliability? Suggestion: don’t buy 2 HDDs from same manufacturer, buy from 2 different manufacturers. When a HDD fails, it’s likely from a defect that will occur in other HDDs from the same manufacturing lot. If you’ve got 2 HDDs from the same lot, your RAID could experience 2 concurrent failures = data loss.

   Reply

5. 3 x 28” monitors for me. 1 central focus (active work tools) + 2 periphery (1 email/web, 1 source documents) neatly matches and delineates my workflow.

   Reply

6. That is a lot of DRAM and disc space. If you want to do AI, it is the DRAM on the video card that matters. I would rather have an external disc server like Synology, or store data in a cloud service, with a fast internet connection.
   The UPS can be external. You can get a $10 usb reader for sd cards.

   Reply

   • I need the RAM for Chrome! If present trends continue it will be 2 GB per open tab soon. I have an external UPS right now to deal with the occasional thunderstorm power glitch. I’d like to clear up my desktop.

   • The politically incorrect Brave browser reports 79.1MB for the philip.greenspun.com/blog page.

     The commenting system here doesn’t seem to use JS, because you can post with a text-only browser like links. That saves a lot of RAM. 🙂

   • Yeah I thought maximising video ram if you’re training / using LLMs locally.
     You can do better than 16gb. RTX6000 has 48Gb but at $12k+ you’d have to be a real money bags to buy it.

7. My PC is an old I5-3570 and runs windows 11-pro without any problem. The only advice is tu bien the iso in. USB stick with Rufus and the proper settings

   Reply

   • Mark: Thanks. I’ve heard that it can be possible to install Windows 11 on hardware that doesn’t meet the official requirements. However, I use this machine for 1500+ hours per year (a depressing thought! So many hours sitting at a desk). It will be valuable to me if everything becomes more responsive (right now File Explorer can be quite slow, e.g., for renaming a file).

8. Nobody has any ideas for a case?

   Reply

   • I’ve got an email to my brother, who knows a lot about custom PC builds. He’s hard to get a hold of, especially by me. If he gets back to me before you’ve made a decision, I’ll post his ideas.

     Have you tried NewEgg’s customer service to recommend a case? I mean you are spending a lot there, probably all AI, but you never know. https://kb.newegg.com/contact-us Might also be a retail PC builder near you, always good to support the locals.

     I can’t remember who offhand, but there are some modder fanatics on YouTube who review high-end cases. Probably a lot of goo to sort through with that route. 🙂

   • HTH. I bragged that PhilG sometimes reads my comments, and my brother had the following comments almost instantly (he’s not affiliated with any of these):

     —

     I’d start here:

     https://pcpartpicker.com/products/case/

     (Phanteks makes good cases, I really like mine. Fractal Design cases are pretty good too.)

     The Cooler Master cooler may be a little weak for the CPU, I personally wouldn’t use it to cool the listed CPU.

     This one is only $7 more and is a monster:

     https://www.amazon.com/Thermalright-Peerless-Assassin-Technology-2000RPM/dp/B0DRFQDSFS

     [Product name sounds badder-ass too.]

     (I’d check it over really carefully though, I got one that I had to send back because quality control didn’t check it out carefully enough, came with wrong Intel bracket.)

     Sounds like that is going to be a pretty sweet PC.

     [He’s apparently a Cartman fan, too, at the age of 60.]

   • Fractal has list of cases that have 5.25″ bays on their website: https://support.fractal-design.com/support/solutions/articles/4000159071-what-cases-support-odd-s-or-other-5-25-devices.

     Facing the same problem regarding 5.25″ drive bays for my next build (but do I really need an internal 5.25″ bay for an optical drive I barely use?). Fractal makes the Pop Air which reviewers say has decent airflow and that two 5.25″ bays (unfortunately they’re located at bottom) – https://www.fractal-design.com/products/cases/pop/pop-air/black-solid/.

     I’ve been using an ancient Antec Sonata case, which has terrible airflow but includes not only 2 5.25″ bays but also a 3.5″ bay .

9. Interestingly, according to Tom’s, for a fee, Microsoft is offering security in Win 10 for a year:

   https://learn.microsoft.com/en-us/windows/whats-new/extended-security-updates

   > For individuals or Windows 10 Home customers, Extended Security Updates for Windows 10 will be available for purchase at $30 for one year.

   Reply

10. +1 for the phanteks cases. I got the Phanteks Enthoo Pro 2 Server Edition once and the only bad thing about it is that it is huge. Otherwise, it can hold as many hard drives, fans, GPU etc as you like.

    Reply

    • But I already have “huge”! I don’t want to overdo the next one. The question is how big a case do I need in order to get reasonably quiet and maybe to hold four 3.5″ drives (two actual 3.5″ drives immediately, room for one extra just in case, and maybe use a 3.5-inch slot in the front if one day the dream of a built-in SD/CFexpress reader comes true and available in a 3.5″ format; here’s a crummy one: https://www.officedepot.com/a/products/338258/StarTechcom-35in-Front-Bay-22-in/ ).

    • @phil, I’m seriously, PC Part Picker is your friend. You can sort and filter. They also have a forum, if you are willing to go into the Internet trenches like us. They have dimensions and volume for each case model. I think they make mid-size cases with 4+ 3.5 bays. AA’s Server Edition is a full sized tower.

    • AT: Fractal calls the Define 7 that I have a “mid-tower” case. I call it “huge”. The forums idea seems like a good one. I’m sure that there are some pitfalls even if a case says that it supports N bays.

    • @philg

      Maybe narrow it down on PCPP with Filters, and check for PCPP, Amazon, and other reviews. You can ignore the Type, and just sort by the External volume:

      Filtered by:Manufacturer: Fractal Design Phanteks

      + Phanteks Enthoo Pro 2 Server Edition
      78.0 liters external volume
      + Fractal Design Define 7
      62.4 liters external volume
      + Fractal Design Define 7 XL
      82.0 liters external volume

      Is yours the 7 or 7 XL? Huge? Try this one, the Ford Excursion of cases:
      + Phanteks Enthoo Elite
      124.5 liters external volume

      maybe you could incorporate it into your interior design, a statement of excess. It weighs about 100 lbs. 🙂 The forums might be able to help with your build as a whole, or other components.

    • Mine is the standard Define 7. It is too big to fit into a standard under-desk PC bracket. A full-size laptop computer is about 2 liters of volume so it is more than 30X the volume of a laptop that mostly does the same job (single 2.5″ HDD for an old-school laptop). Maybe the answer to making a compact case work is abandoning the 3.5″ mechanical drives. Just have two front slots for them (in case the dream of a card reader comes true one day), but rely on 2.5″ mechanical or SSDs. That would enable the use of the Define 7 Compact (two 3.5″ bays and room for up to four 2.5″ disks). I probably could start out with just one big 3.5″ disk drive as a versioning backup. Maybe add a second 8 TB M.2 SSD for working with big files occasionally.

11. I have an AMD 5900. I have a Noctua D15 cooler. It’s great and super quiet.

    https://www.youtube.com/watch?v=79sbl1HPFcU

    Heatsink on m2s: some motherboards will have a gigantic (surface area) heatsink that is used for m2s and the chipset. If your board had this you wouldn’t need heatsinks on the m2s making use of it.

    Also re: m2s, open up the motherboard manual and read specifically which m2 ports supper which versions. There will be CPU and chipset managed drives and max versions. You want to make sure for ever v5 m2 you have an actual v5 m2 slot you use and not a v4 (otherwise you’re wasting money).

    You might be better off with your idea of all ssds for workflow and 7200 RPM just for backups. Modern apps (and their developers) often just assume ssds. My machine is from 2021, so I have one 2TB m2 for the OS and programs and another one for “data”. You could do the same with 2 8TB m2’s if you were trying to maximize throughput when doing intensive tasks.

    You mention choosing the motherboard for extra USBC on the back. This again is something you have to double check exactly what speed/version of USBC you want it to be. Some motherboards might have headers for more. From previous research, there seem to be external cards you can add to get 4+ more. Something to consider.

    Do you have a specific idea of what video card you’d need for the local AI search? I’m concerned the standard 32gb ram in 5090’s or some other 5090 only quality might be more AI preferable. However, besides that modern cards are insanely fast for everything else you’re doing. That said, if you’re not immediately going to do the AI, you can just get the 5080 and if isn’t enough you can get a 6x or 7x line whenever you do the AI stuff. I would say though, if you’re not playing games I’d expect a 3080 to run 2 4k’s fine. You might see the difference if you were encoding lots of video or doing lots of photoshop stuff, but I’m unsure. I’m not saying this because I’m suggesting saving money on a lower card, but rather that given your stated application use I see only the AI part as an issue.

    I did a bit of googling and if we wish to believe the results this article on the subject seems very useful:

    https://www.pugetsystems.com/labs/articles/nvidia-geforce-rtx-5090-amp-5080-ai-review/

    I ended up buying a Graphics Card GPU Brace Support, Video Card Sag Holder Bracket, GPU Stand (L, 74-120mm) to keep my 3080 from not sagging. It’s a bit laughable, but a lot of these modern cards just are so heavy.

    https://www.amazon.com/dp/B09FPJL1KY?th=1

    re: Case, I have a Be Quiet 500 case (I think an older version). It’s very nice to work with and quiet. I have no complaints other than I wish there were more USB ports on the front.

    I have a 5900 12core/24thread w/ 64gb ram w/ 3080 and the same general workload as you and everything is super super fast. I routinely run 500 tabs in chrome without any problems or any slowdown. If you weren’t playing games full screen the video card wouldn’t change anything. If you wanted to play video games full screen you’d want the 5080.

    (long time reader since the Panda book!, feel free to email if you have personal questions)

    Reply

    • Thanks, Michael. I didn’t realize that the heat sink on the M.2 drive might be redundant. I will have to study motherboards more carefully. Thanks for the Puget Systems link. We’re informed that every day and in every way electronics get better and better. Yet, the AI TOPS numbers for the brand new (and exciting!) 5080 are barely superior to the 4080 (2022). It doesn’t look as though Moore’s Law is working in GPUs or in RAM or in anything else that I need to buy. I guess the other thing interesting from that chart is that the 5090 is almost twice as powerful as the 5080 for 2X the price. I would have expected more dramatically diminishing returns.

    • It looks like you’re right about the heat sink. https://www.asrock.com/mb/AMD/X870%20LiveMixer%20WiFi/index.asp shows a “toolless multi-layer M.2 Heatsink” that is part of the motherboard. I can’t quite figure out from the photo how many M.2 drives it covers.

      Now that I look at the fine print on the motherboard it seems that actually using the M.2 slots reduces USB capability: “If M2_2 is occupied, both rear USB4 Type-C Ports and M2_2 will downgrade to x2 mode. You can switch M2_2 to x4 mode in BIOS setting, but doing so will disable USB4_TC1 and USB4_TC2.
      If M2_3 is occupied, PCIE2 will downgrade to x2 mode.
      If M2_4 is occupied, PCIE3 will downgrade to x2 mode.”

      (i.e., you get what you expect from this motherboard only if there is no more than one M.2 drive)

    • It must be a CPU/chipset issue because the expensive motherboard has all of the same limitations regarding what happens if you have the temerity to try to use the M.2 slots. https://www.asrock.com/mb/AMD/X870%20Taichi%20Creator/index.asp#Specification

      It is too bad that my 16 TB M.2 dream can’t be a reality 🙁

      (I guess putting the second drive in M2_4 wouldn’t be super toxic since it mostly trashes a PCI slot that I wouldn’t use anyway.)

12. I went with a home theatre case last time I did a build. Not too big at all, about 26 litres. A video card fit inside, along with a card reader, hard drive swap slot and DVD on the front. I recall I might have had to choose a small form factor power supply.

    Although my wife still thought it huge. Ha.

    Not sure if Liam Li still sells these PC-V800

    https://www.newegg.com/lian-li-atx-desktop-aluminum-computer-case-black-pc-v800b/p/N82E16811112082?srsltid=AfmBOooRHFdMWNFkCh61v82ESgjj1pukS-oqZQCVPg-GyGRVnJlOMehW

    Reply

13. If you are planning to train AI models locally you need a decent GPU, either older RTX 4090 or newer RTX 5090. They both require power and specialize cooling. I would honestly just buy already prebuilt box, with AMD if you so fancy. CPU isn’t that important, the GPUs VRAM is what drives these systems

    I assume you know how to load up your local llm into GPU and tune things up etc

    Reply

    • Thanks, Tim. It might be too ambitious to train with 16 GB. Maybe the goal should be scaled back to training using cloud servers and just being able to, maybe, run a model locally. Perhaps the whole idea of a desktop computer and local computation no longer makes sense. In the 1960s we had time-sharing right. We paid for the computer during the few hours per week when we needed it. AI now requires enormous resources. Maybe $50,000 in hardware to train anything and $10,000 of hardware to run a model? At the same time, the individual needs to use this hardware only for a few minutes or hours per week. So the right thing to do is pay for fiber Internet (I would kill for that opportunity! Stuck in an Xfinity cable ghetto here) and get a $1,500 laptop plus nice monitor and keyboard and use them to connect to cloud-based services.

14. Speaking of monster-sized cases: Cantrill might give you a discount on one of these bad boys, maybe sell off one of your copters:

    https://vmblog.com/images/oxide-team.jpg

    They are hiring (some remote), I think the salary is ~$200K plus equity.

    https://oxide.computer/careers

    Reply

15. > I guess the other thing interesting from that chart is that the 5090 is almost twice as powerful as the 5080 for 2X the price. I would have expected more dramatically diminishing returns.

    NVidia engineers it like that on purpose to finely manage (ie force a tiered pricing model on) their perceived market segments. BTW if by chance you missed the March keynote, and otherwise haven’t been following super closely it is truly amazing what they are doing. (I’m really not a keynote kind of person either)

    GTC March 2025 Keynote with NVIDIA CEO Jensen Huang
    https://www.youtube.com/watch?v=_waPvOwL9Z8

    > Re: M2 Heatsink

    Go to your link and browser search “PCIe Gen5 Blazing M.2” and there is a photo showing it. The keep talking about it if you scroll down from that photo.

    > Re: M2 downgrading PCI

    Note that M2_4 only supports Gen4x4 instead of the Gen5x4 of M2_1 and M2_2. The speed is already so insane on a 4×4 and unclear if outside of benchmarks you’d ever see the difference that this might be completely acceptable for the second drive. If I were in your situation I’d probably take this option. A bit more info:

    More importantly: Gen5x2 == Gen4x4 == ~8000MB/s. Gen5x4 == ~15000MB/s.

    If you were to use M2_4 your Gen5x4 SSD will downgrade to a Gen4x4.

    It looks like the USB4 ports are USB4 v1 limited to 40 Gbps

    If you really wanted it, you could look for a board that supported USB4 v2 which does 80 Gbps or even 120 Gbps. I’m not sure they are generally available yet though.

    > Re: CPU/Chipset

    Yes, we’re a far cry away from the pack-in-8-ISA-cards-and-hope-for-the-best days. If you want to see an answer for how some people do achieve this google “motherboard with 4 gen5x4 m2s” and read the Gemini answer. One take away is the existence of: ASUS Hyper M.2 x16 Gen5 Card

    > Re: Not training locally

    I know people training locally with this kind of hardware and they have interesting results. There’s all kinds of things you can do. Unless you had unrealistic goals I wouldn’t stop pursuing it. NVidia also has special desktop hardware they sell for home and devs and new graphics cards will come out every so often. Also: cloud GPUs are not cheap at all.

    > Re: Oxide

    Yes, Oxide most definitely needs to sponsor Philip’s next project.

    Reply

16. silverstone makes nice looking cases with good schematic documentation and external drive bays.

    Except for the AI training (i.e. actual computing with big numbers), a good enterprise or workstation laptop will suit your needs and travel well. When at home, you can hook it up to a nice keyboard and monitor.

    HP and LG make nice machines. Touchscreens are great, especially when doing graphics work. The LG Gram series are beautifully light. HP has the best documentation and engineering. You absolutely do not need a fancy grsphics card unless you are gaming.

    As for the AI machine, think like you are building a server, not a user terminal. Older parts can be very cost effective for the vector computing components (grsphics cards). It is my understanding that AI is all about how much RAM you have. If you build sn ai model that works well with a traditional CPU then RAM constraints are much less.

    Reply