The custom APU at the core of Sony’s PlayStation 5 hasn’t just been quietly powering these game consoles, but also made their way onto cryptomining cards around 2023 which are called the BC-250. The APUs on these boards differ from the one found in the PS5 most notably by having two out of eight CPU cores disabled, along with many compute units (CUs) of the iGPU. Now apparently it seems that you can re-enable these CUs per instructions by [duggasco] if you’re feeling adventurous.
As stated in the project’s README, BC-250 boards come with only 24 out of 40 CUs enabled, but this is not a permanent (e-fuse) thing. Instead you can write to two hardware registers during the GPU driver initialization, something which can be added to for example the Linux kernel module parameters.
Since many of these APUs likely had cores and CUs disabled due to them failing QA during PS5 APU manufacturing, there’s a good chance that some of the CUs truly are bad. Yet as we saw with the AMD Phenom II X3 with a supposedly bad fourth core back in the day, sometimes demand for the ‘defective’ part is high enough that good parts get mixed in as well.
Thus people like [Lowest Logan] decided to give it a shot, demonstrating the use of the patch with Bazzite Linux on a BC-250 system. After a reboot the system does indeed list 40 CUs as being enabled, and running Furmark shows a big boost in performance without any glitches or fire. There is of course thermal throttling, but that is due to the default cooling solution not being designed for running it at full blast.
Incidentally the real PS5 has only 36 active CUs, so this technically makes these unlocked APUs more powerful. With the water cooling solution demonstrated by [Lowest Logan] the thermal throttling is also resolved, showing that you can get a pretty nice gaming system out of these old cryptomining boards if you happen to win the silicon lottery.
Interestingly (to me) was spotting a few days ago that these BC-250 boards are being repurposed for local AI inference, this bloke has it pinned on his xitter https://x.com/loktar00 – I had thought this claim had a bit of a smell about it, but maybe not.
“This is actually CRAZY!!! Using llama.cpp RPC
I have 2 BC-250’s setup so far, they’re able to run Qwen 27b at Q4, and 35b at Q4 as well.
This is without extra CUs unlocked:
Qwen 27b with MTP – 14.5 tk/s
Qwen 35b with MTP – 47 tk/s
For $300 I’m getting these speeds! This is wild!”
I assumed it was because people were building their own “steam machines” with these, as they are roughly the same performance and size and it’s possible to put SteamOS on these.
I think they only have 16GB of RAM for cpu and gpu which is a little stingy for AI.
Shame the price of these things have been quite inflated for quite some time. I definitely missed the boat. I’m mildly jealous.
Same. And it appears to me it’s a US only model. I was trying to find some locally in the EU but I’m unable to find them. I wish I bought a bunch of them when they were around 80 USD on eBay. They appear to be quite powerful.
Lets ignore the chinese webshops? Sadly i never seen them for below 130 on there myself but with some discounts and coins, one can reach tempting prices. Upcomming 3 euro extra import tax made me just order one and i hope to see mine at the end of june. For gaming and ai 🤩
Bought one last week there for 130$
Quite more expensive but still awesome value
Wouldn’t be surprised if they were disabled simply to improve the perf-per-watt ratio or something. I wouldn’t be brave enough to try enabling them though.
Enabling the cores greatly improves the perf-per-watt ratio. The same furmark score with 40CU can be achieved at much lower GPU clock than with 24CU, at around 100W less power from the wall.
As far as I know the activated CUs are also overclocked. So I assume these boards where optimized for workloads that require core speed more than parallelism.
No, all CU use the same clock speed.
Reminds me of the good ‘ol Celeron 300A days
Be nice to run some SteamOS with some work naturally.
It works ok most run bazzite or cachyos with game mode. there are some running real steamos also but more hassel to modifying the kernel and getting all govenors working for OC.
CachyOS with their shipped steam enhancement.
I assume the clock speed can be adjusted — sometimes when AMD disabled CPU cores (to sell as a lower SKU) rather than them being actually defective, they just disabled the ones that had the lowest top boost speed. So I do wonder if in some cases one couldn’t enable all 40, find it’s unstable, and just shave a few % off the clock speed to have it fully stable. Anyway, very cool!
I made a PR and added docs here weeks ago https://elektricm.github.io/amd-bc250-docs/linux/fedora-coreos/
I’m clustering 3x of these and running Gemma4 26b, Gemma4 12b (with audio/image encoderless), etc. They’re working fairly well;
Gemma4 E4B -> 800 tks prefill, 60 tks Gen
Gemma4 12b -> 300 tks prefill, 40 tks Gen
RPC Clustered Gemma4 26b -> 300 tks prefill, 30 tks Gen
Be careful on heat generation here.