Thread 106890879

>/g/ - Technology
>no thread about DGX Spark release
Are you serious guys?? I barely browse /g/, why do I have to be the one to bring it up...

Hobbyist generative AI platform from NVIDIA in a small form factor.
MSRP $4000 gets you onto the DGX ecosystem (/utwg/ are you listening??? DGX is what enterprises use for machine learning, hint hint).
Already set up and ready to go for local models (/lmg/ and /sdg/ are you listening?? prompt gooning has never been easier and more liberated)
Can link two of them together for increased VRAM and performance (cool, and also means if you regret your purchase, someone will be looking to turn 1 Spark into a 2 Spark cluster, making it easy to sell and get your money back).

Reviews:
https://www.youtube.com/watch?v=Lqd2EuJwOuw
https://www.youtube.com/watch?v=FYL9e_aqZY0
https://www.youtube.com/watch?v=zs-J9sKxvoM

Are you getting one? Are you planning on a Mac Studio instead?

buy an ad

Sounds great until 2 years pass and its on ebay for a couple hundred bucks.

>>106890879 (OP)
I may be interested indeed, but I am poor.

(That notch in the screen... That bloody fucking retarded notch in the screen...)

>>106890879 (OP)
looking forward to buying this for $50 from the local recycler in 2 or 3 years

>>106890879 (OP)
Kek this costs more $ than my display has pixels. Thanks, I'll wait.

>>106890952
You have a 64x64 display? Even 720p is almost a million pixels.

>>106890879 (OP)
No, I don't need more.

>>106890879 (OP)
slow RAM. nvidia always fucks something up.

>>106890879 (OP)
nice try jensen
but I already have 128gb of cheap ram

>>106890879 (OP)
$4000 to get a system that will just become e-waste in 2 years as people dump it for the Next New Thing?

Just get a Strix Halo system and throw 128GB on it. It can do inference just fine with most open models, and if you're fine-tuning something you can just rent an instance. Unless you're going to be training something 24/7 for months at a time it will be cheaper.

>>106891734
I'm sure one day when the price goes way down they'll be interesting, but it's just not competitive when you compare it to strix halo. I think this generation is just a stepping stone anyway. These all really need 256GB. The only thing I've picked up on from strix halo and the spark as that there's likely a spot in the homelab alongside NAS and such for an AI box.

>>106890879 (OP)
>/utwg/ are you listening???
Bro nobody unemployed has 4000 mcdonalds coupons to piss away on a meme slop appliance.

>>106890900
The way Chine is doing trade war shit with their processed minerals, these actually might go up in price.

>>106890879 (OP)
what do

>128GB ram
>$4k
I don't care how many tflops this bitch has I'm sticking to strix halo. AI isn't good enough yet to justify this.

>>106892036
>AI isn't good enough yet to justify this
that's what they're hoping you don't nootice

>>106890879 (OP)
>Already set up and ready to go for local models (/lmg/ and /sdg/ are you listening?? prompt gooning has never been easier and more liberated)
It has around half the performance that the Strix Halo does for LLMs, costs twice as much, and uses twice as much power. It also has the same memory (shared by the way, slightly faster at 273 GB/s instead of 256 GB/s) and runs on Arm, so it has worse software support.
>CUDA
Yeah, CUDA. It still has worse software support because it's Arm instead of x86. I'd get one if it was 2k, but at 4k it's just fucking egregiously overpriced.

>>106890879 (OP)
project digits sounded like a game changer when announced now it's extremely overpriced and barely competitive with applel and aymdd

>>106892116
It was obvious it was going to be fucking dogshit when they refused to give any specifications about it.

>>106890879 (OP)
can i use it as a regular pc

>>106890879 (OP)
I'm getting shilled this garbage and as usual it looks to be a slightly better but massively more overpriced version of what AMD already produces.
Fuck Nvidia and their b2b shift. Abandoned the consumers.