In Switzerland the 64 GB configuration of the DIY edition desktop is out of stock but the mainboard alone is not.
Yea the RDNA 3.5 40 CU is too slow for generation unless most of the computation is in 4 bit, but itās still not the same experience as an API where youāre getting 4000t/s prefill.
You can get 1000-1500 t/s prefill on qwen 3.8 flash next with Gufo: the all-in-one strix halo inference engine or Halogen; Iāve tried both locally and they are a real thing. At this rate agentic coding is usable; but itād need another 3x to not feel sluggish.
No kidneys left. Already running on external power.
2 Likes
ā¦what Alex_Morozov said (and others similar).
As a hobbyist Iāll stick with the 128Gigs model until a RDNA4 APU is shipped with 192 or 256Gig HBM support. The 495 would be perfect for a 1st purchase. Not worth it for existing 395/128 owners unless one absolutely needs more RAM.
In that regard AMD is about to be left behind again. As NVDIA has inked relationships with both ARM and Intel to deliver CPU+GPU APUs (flush the NPUā¦ Iāve used the Lemonade LLMs on the NPU and I question why bother. Iām plugged in to the wallā¦ why do I care about a few watts saved at the cost of much slower performance). If NVDIA has a similar offering as AMD with RDNA4 Iām still going the NVIDIA product next time.
I wish I bought the 128GB desktop last year when I was looking at starting up my homelab. I ended up going with a AceMagic 370 instead not realizing how much weaker its GPU is.
Now Iām considering between the new Mac Mini and the 64GB Desktop (assuming it ever comes back in stock), but itās hard to justify a $3,200 CAD purchase for what would effectively be an offload system for me at this point.
Framework, if you release a 48GB version of the AI Max 495 as a mainboard-only option Iād buy it instantly.
Medusa Halo is the next one, and rumors are that it skips straight over RDNA4 and goes to RDNA5, 48CU. 384bit LPDDR6. Up to 24 high performance cores (and maybe a couple tiny low power ones.) The equivalent of the 385 would be 12+2 and 28CU from the sound of it.
Like you said, as a hobbyist, thereās not that much useful in the Gorgon Halo chip. I am looking forward to see what Framework does with Medusa Halo though. Probably at least a year off though. Even though I only have a 64 gig variant of the old board, I still donāt have a need for the Gorgon board. I hope they release a 96 gig variant of Medusa though.
They would not go with a 384 bit wide bus. Thatās more die space they wonāt have as bigger players are the most advance silicon nodes locked down.
Well, OK. Only mentioning what the leaks were. Iād be surprised if they couldnāt do 384 on a reasonably sized APU on a N3 node though, since xbox series x managed a 320 bit on a 6nm node.
The rumors are that there will be a version with 256bit bus as well, that uses the existing socket and LPDDR5 still, so I wouldnāt be surprised if the first versions released are just the same AI495+ board we have now but with the new die slapped on it, and the same bus.
The timestamp linked is where heās showing the bit about the bus width there. Itās the latest leak from June.
Interesting video. It would be more believable that Medusa Halo with the 384 bit wide bus would be the next gen xbox / playstation APUās and the Medusa Halo miniās with the 128 bit bus would be āAI boxā that framework would sell.
If we do the calculation y =ax+b where y is the final price a is dollar per GB per RAM, x is the amount of RAM and b is everything except the RAM. Is b negative again?
I wouldnāt even bother with the Medusa Halo Mini, itās not really an upgrade from the Strix Halo, just like the Gorgon Halo is barely an upgrade. The Halo Mini might end up in framework laptops though, given its much smaller consumption.