>Ryzen AI MAX+ 395
>RADEON 8090S iGPU
>128GB of shared LPDDR5X memory
>can run gpt-oss-120b super fast
>can run Qwen3 235B 4bit
>costs $2000
Why don't you own one? You do want to run a 235B model locally without government interference, right? You aren't poor, right?

>>106216635 (OP)
>without government interference, right?
>chinese PC

>>106216635 (OP)
i remember reading somewhere the ai inference slower than a 2060. also you cant run 235B 4bit on it. i think only about 90gb are usable rest are reserved

>>106216635 (OP)
>sends your data to Beijing every 1.6 attoseconds
No thanks

>>106216670
what makes a pc chinese?

>>106216635 (OP)
128 isn't enough. if this is supposed to be an ai computer it really needs 256.

>>106216743
Being sold by Chinese owned company

>>106216732
>*disconnects it from the internet*
what now?

>>106216756
so like your favorite lenovo stinkpads?

>>106216763
>disconnects
>implying

>>106216682
It can run the Q3 qwen-235b at 13 tokens per second and the Q4 at around 1.5 tokens per second. Yes Q4 is slow but still useable with telling it to generate shorter answers.

>>106216778
what, it cracks its nearest wi-fi password through pure i-ching seethe and silently connects and pings?

>>106216822
>What is GSM

>>106216776
lmao

>>106216776
I prefer HP

>>106216776
you can't just drop a trvke like that on /g/ sir

>>106216745
It costs $1,999 for a full blown PC with 128GB of unified memory. Nvidia gives you 96GB for $7000 and that's just the GPU.

>>106217270
i expected this. if you really want to run large models you put certain things on the gpu and overflow onto the cpu for less used stuff. this thing only has one memory pool, so it needs more. sure it can run larger models, but if you want something like the real thing you're gonna need more ram. this thing falls into the gap between "playing around for fun" and "useful for real work" because of ram limits.

>>106216635 (OP)
Mobile cpu is a stopgap. You can get 4 AMD Mi50 cards with 32gb vram for $200 each. $800 plus another $1200 for an epyc build and you can host the whole model in vram with like 4x the bandwidth. Vulkan is a little slower than native cuda but you can easily get 30t/s instead of the <10 you'd have on a mobile build.
I wouldn't bother with rocm.

>>106216635 (OP)
what can I do with it besides asking it to estimate the calories in food and generate porn? seems kind of gimmicky

>>106218003
Anything, it's still a full blown miniPC. The Radeon 8060S is even half decent at gaming, around RTX 4060 level.

>>106216808
>13 tokens per second and the
> Q4 at around 1.5 tokens per second
Lmao lmao lmao

>>106218066
https://kyuz0.github.io/amd-strix-halo-toolboxes/

>>106216732
glowie hands typed this

How well would this thing work for local image generation?

>>106216952
>what is a faraday cage

>>106216635 (OP)
>$2000
>for chinkshit support and horrible cooling
The HP one is 2500 on sale, that 500 dollar difference is worth it to avoid the chinesium.

AI is fucking retarded, OP is a faggot.

>>106218470
not even subtle

>>106218589
?

>>106216635 (OP)
the bandwidth is too low

i'm so tired of /g/ hostility toward AI. you idiots should be the most hyped by far. this is THE MOMENT, it's here and you all cry like xitter cartoonists. it's pathetic.

put them in a credible 16-18" workstation laptop and I'll consider it.

>2x faster then normal CPU inference due to doubled memory bandwidth.
I mean it's not bad but it's not great either.

>>106216635 (OP)
My current PC runs circles around it. Why would I?

Trash. This isn't unified ram, so the inference is slower than dogshit

Thanks but I'll wait for Nova Lake AX

>>106219196
I've lost some hype some post-GPT-5, seems like we're only going to get marginal upgrades until some new breakthrough.

>quad channel lpddr5
isn't DDR6 due out soon which should double speeds anyway?

>>106216635 (OP)
>AMD

>>106219409
This is great news, because it implies the technology can finally trickle down to the edge. GPU prices are finally starting to stabilize