Join the discussion

Write your take first — we'll ask for email only when you're ready to publish.

  • Hacker News
  • These are going for a few hundred dollars now due to demand.
  • For anyone wondering about local LLM applications:

    - current price $200+

    - limited VRAM (12-14 GB)

    - poor performamce in clusters [1] - only 2 lanes of PCI-e 2.0 (still better than CMP XX HX mining cards)

    [1] https://ciroluciotecce.it/en/posts/bc250-cluster-inferenza-l...

  • Jeez I checked eBay and they are up to $300+ now
  • This isn't necessary related to this project but I got thinking about while reading it. Isn't it weird that shared memory was a cost saving measure by not having GPU dedicated has now been rechristened unified memory and is now a feature?

    I understand why the AI people want it to help with the latency between the CPU and GPU. But having people in some cases use it as for marketing is kinda like getting watered down beer and getting told its a good thing because alcohol is bad for you.

  • > want it to help with the latency

    Primarily the bottleneck actually. PCIe bandwidth is extremely limited.

    GDDR has worse latency than DDR and scaling it up to larger capacities faces challenges. So if you unify memory your options are small GDDR, large DDR, or expensive HBM.

    Unusually, the BC-250 and PS5 both use GDDR with the CPU so the memory bandwidth is great but the latency not so much AFAIK. It's like using a dGPU as a full computer.

  • > I understand why the AI people want it to help with the latency between the CPU and GPU.

    The reason AI users like it is because it provides a large amount of higher-bandwidth memory at a price much lower than an equivalent GPU.

    If you had 16GB of DDR5 in a normal 2-channel consumer board it would not be this fast.

    This gives 16GB of high speed memory for a very low price compared to a full GPU.

    The larger unified memory systems like Strix Halo or Apple Silicon provide access to high bandwidth memory at scales you can’t get with any consumer GPU.

    > But having people in some cases use it as for marketing is kinda like getting watered down beer and getting told its a good thing because alcohol is bad for you

    I don’t think you understand the benefits of having high speed memory attached at a low price point.

    Yes we would all prefer to have the same amount of even faster memory attached to a dedicated GPU, but that’s not happening at this price point.

    Better analogy would be that this is an entry-level sports car for the price of a Honda Civic but you’re complaining that it’s not a Lamborghini.

  • Technically, “unified memory” is where both the GPU and CPU have access to the RAM over the same bus, which is different from the older, reviled “shared memory”, where the RAM was always at least one bus hop away from the GPU. But yes, there is a certain amount of spin to it, isn't there? (“You have less total memory now, and that's a Good Thing!”)
  • "many PS5 APUs that didn’t meet console spec (“binned” or “defective” for console use) got repurposed. Instead of being discarded, these chips were re-used for mining/higher-density hardware."

    Not discussed much in the article, but it looks like:

    TSMC → AMD controlled/qualified the usable off-spec parts → AMD supplied BC-250 parts to ASRock → ASRock sold mining hardware.

  • I think this is one of the more interesting parts of this machine. How did they end up at ASRock?
  • Here's the GitHub repository linked by the (currently unreachable) article in the OP:

    https://github.com/csabakecskemeti/amd_bc-250_how-to

  • There's a similar board called the AMD 4800S (or 4700S). These are also repurposed PS5 SoC that didn't pass the graphics quality control. So what you get is a motherboard with CPU and 16GB RAM soldered on. No HDMI out, but it has a 16x PCIe slot (PCIe 4x 4.0) to connect a graphics card. The 4700S model is limited to 4x PCIe 2.0.
  • I checked it out and total cost is too high compared to what you get in modern Mini PCs. Also, the 4800S is difficult to get and 4700S supports only SATA SSDs.
  • Does anyone else immediately lose interest in an article when you realize it was written by AI?
  • Is it? Pangram disagrees and human-made photos imply at least some human effort.
  • Yes but I'm also getting a bit tired of reading complaints about it as well.
  • $60 gaming PC's are doable if you buy an "untested" Dell Optiplex and get lucky. A few years ago, I bought a couple of these in that price range that not only worked, but had a GPU of some kind. Good enough for Left4Dead at least, and some other Indie games. All they needed was an HDD with Windows 10 (both of which I had lying around - Win10 licenses can be had for a few bucks).
  • > $60 gaming PC's are doable if you buy an "untested" Dell Optiplex and get lucky.

    Yeah, but a $60 lottery ticket that pays off in an X if you win isn’t a $60 X.

  • As long as you only want to play games from 2008. :)
  • You are _not_ getting one of these for less than $300, and because of this post and many like it, there are countless scams selling just cases for them at this price from a dollar or two of filament.
  • How is it a scam to sell something for more than the cost to make it? Labor and energy are not free you know.
  • If it's advertised as a case and it works, why is it a scam? I don't own a 3D printer, nor do I know anyone who does. If someone is designing, printing, and selling a useful product, why is that a scam?
  • This always makes me ambivalent about enthusiast blog posts. Like on the one hand, cool that the knowledge gets spread. On the other hand, whenever one gets popular, another hidden gem skyrockets in price or becomes semi-permanently low in stock.
  • I bought one while waiting to see if my Steam Machine reservation would be fulfilled.

    I paid ~$186 shipped for the board on eBay 5 weeks ago. I use an ATX power supply and NVMe that I had laying around and I printed a big case for it[0].

    I seem to have gotten a good one as it has no problems after unlocking the extra CPU/GPU cores.

    I run vanilla Arch on mine configured to boot into Steam. The jank factor is pretty high but it does trade blows with the Steam Machine for a fraction of the price.[1]

    I did end up going through with the purchase of a 512GB Steam Machine for $1160 w/tax and swapped in a 2TB NVMe. It is definitely a more polished/living room-ready product and I'm happy with the hardware but it's definitely a questionable value.

    It's worth reading over the docs[2] first before committing. There are lots of options for cooling, storage, and power supplies. Just depends on what you want to get out of it and how much time you want to invest in it. There are pre-made images so you don't necessarily need to be a Linux expert but it would help if you want to stay on top of the latest developments. There is still on-going work to get even more performance[3]

    - [0] https://www.printables.com/model/1789618-bc-250-atx-case-dua...

    - [1] https://www.youtube.com/watch?v=5Sw1SdU9buQ

    - [2] https://elektricm.github.io/amd-bc250-docs/

    - [3] https://github.com/lonewolf0622/BC250-Native-Mesh-Shaders-

  • What's the performance delta? I'm waiting on a SM reservation but I wouldn't mind saving $500+ :)

    Is there any "jank" during active use or is this mostly a set up thing?

  • I built one of these. It is not possible to do so for 60$ anymore since the board itself goes for 150+. On top of that you still need a PSU, NVMe, a high pressure fan, DP to HDMI adapter, possibly BT and Wifi adapters, and a 3d-printed or DIY case.

    It is also a very hacky build. You need to flash the BIOS which enables a few interesting unlocks: from 24 to 40 GPU compute units and 6 to 8 CPU cores. It's a bit of a silicon lottery and will depend on your specific board and you need to test everything thoroughly yourself, honing on a config that works for you.

    The software side of things is Linux, so if you're familiar with setting it up it shouldn't be difficult to go through the different steps. There are a few distros that work well. I use Bazzite and boot directly to Steam big picture, the UX is great. You can exit to desktop and you have a fully usable Linux box that you can play around with from the comfort of your sofa.

    One of the downsides is that it uses a lot of power when idling, about 80W. I just turn it off after using it. I'm working on a QOL feature: an ESP32-based power switch that starts the board when a specific BT device is detected (usually a controller). Once everything is setup the board works very nicely. It can be a bit loud depending on your fan/case combination. I'm currently printing this one[0]

    It delivers about the same punch as a standard PS5. The hardware is remarkably similar, and much better than mini PCs set up with AMD APUs, unless you go Strix halo but that is another price league (10x). You can play Cyberpunk 2077 decently on it on settings similar to an AMD RX 6700XT or slightly below RTX 3070.

    A pretty impressive number is the memory bandwidth at ~410 GB/s, which should run local LLMs like Gemma 4 or Qwen3.8 27B at decent speed but with limited context size as its just 16Gb

    Would I build it again? It was a lot of fun, but financially is not a crazy deal if you need to buy all the materials and factor your time in. Still much cheaper than a Steam Box!

    - [0] https://makerworld.com/en/models/3021754-bc250-case-mkuu-fsp...

    Edit: GPU performance comparison

    by nsbk
  • > you still need ... DP to HDMI adapter

    what for? just use DP

  • Why would the GPU context be limited to only 2GB?
  • Are there DP-HDMI adapters that work with HD audio bitstreaming and VRR? Different use case, but I've not had much luck finding good options.
    by bpye
  • I don’t see how you could run Qwen3.8 27B on 16GB of memory that’s shared with Linux. Are people running models at 2bit quants? Are they even worth bothering with? I had assumed you go down to 4bit and if you need to go smaller you have to lose parameters.