So why don't you?
>>109083124Chinese APIs are just cheaper
>>109083124Because it's a shitty doa device with literally no usecase, you are better off buying a bunch of gpus.
>brave filename>retarded post
>>109083124I can afford it because I don't waste money on things like it.
Because that system is just a bad deal, it's way too slow to be of any real use.It has a memory bandwidth of 273 GB/s. That's almost the same as a 3060.To put this into perspective, 5090 has a bandwidth of 1792 GB/s.But something a lot cheaper like 3090 is at 936 GB/s, and you could buy 4-5 of those for the price of one spark and get way more utility out of them.
>>109083124they arent even good perf wise
>>109083124If I wanted an arm gaming pc I would buy a used macbook with m1 or m2 cpu
>>109083124How much RAM?
>>109083287>It has a memory bandwidth of 273 GB/s. That's almost the same as a 3060.you only need to load the model in to memory once, right?
>>109083381No. You need to work with it in memory too.
>>109083124>Personal AI Desktopuse case?
>>109083381>>109083418BWAHAHAHAHA 500 "people" fell for this shit
>>109083124Wha is this if not the biggest scam they've pulled to date.This is way too slow to be useful to anyone.
>>109083124>arm>can only emulate x86_64 up to sse instructions because of toxic open source communists who shoot down issues and patches relating to adding avx/avx2/avx512 support, the same kind of awful people responsible for gnome being stuck at 60fps with stutters on high refresh rate monitors until 2019-2021 and kwin's still ongoing x11 shittiness that never got properly fixed even with forks(ofc the fork's change got shot down by the same communists with justifications like "lol xd my thinkpad can't show more than 60fps anway and the shitty tn response time hides stutters and tearing")>literally stuck in 2010 pre 2500k era software support(see above)>128bit SIMD only, literal joke, Intel's SSE did it before the start of the milleniumIf you buy this you're an idiot unless you only intend to use this for LLMs or whatever other workloads need 128GB of RAM, it's useless as a desktop PC.
>>109083287You can compensate the bad bandwidth by only using MoE LLMs, but you can already do that with a 5090 or whatever else gpu has lots of vram, and knuckles(normal models)
>>109083432500 "people" fell for this shit in the last month
>>109083477>avx/avx2/avx512 supportpretty sure those have patents on them.
>>109083124I can, but I built a PC in 2023 for like 3.5k that does most everything I want.
I can't afford it :(
>>109083124these spark and helios mini units are dog shit when used for single or low parallel, the tk/s are worse than a GPU of similar cost. they only make sense if you have 15+ users and even then they're slow as shit
So that's why the ram prices are so high, they just want to sell their overpriced cuck boxes to the brainless llm zombies, I'd rather buy an ounce of gold
>>109083287>But something a lot cheaper like 3090 is at 936 GB/s, and you could buy 4-5 of those for the price of one sparkyou cannot buy a 3090 that cheap unless you are throwing in some expert level BJs, ask me how I know
>AI>super computerFeces Buffet
>>109083124Call me when they have something capable of serving a 1T at 1,000 tokens per second to at least 2 users. Don't make this thread again until that is the case. MAKE NO MISTAKES.
>>109083124the lack of airflow is susbut I guess my MacBook m3 pro dgaf about it and works
>>109083321>currybook
>>109083124I already have unfettered access to two nodes with eight H200s.If I need more I can also run jobs on a HPC which has >60 H200s and a bunch of B200 nodes getting installed as we speak.Why would I want to downgrade?
>>109083124>$5k for a very specialized, locked down, machineno thanks, you'd be much better off renting a server/gpu as needed.
>>109083124I know everyone went insane the last few years and deluded themselves into the belief that software is NVIDIA’s moat but historically and to this day NVIDIA has had shit quality software. I like the idea of this device but don’t want to deal with all the software issues and instability.
>>109083124rtx spark is coming out later on in the year
>>109083124I'd rather just buy a blackwell card, thanks.
>>109083124the more they try to shoe horn this AI thing, the less willing I am to open my wallet.
>>109083124Use case?
>>109083124I have no use for one.
>>109084377it doesnt matter. its actually hilarious how little it matters that you dont care.
>>109084393Genning stuff for gooning while you game on your other computer, DGX = Da GoonboX
>>109083124I want a single general use machine and ARM Linux ain't itIt's either an RTX 6000 Pro or an AI Ryzen Max for me.
>>109084770Counterpoint: type "anime titties" into a search engine.
>>109083124Becaue Nvidia basically told us >>109083166, when we wanted to purchase some at work. We have direct sales contact point because we are basically a dc on enterprise scale, yet they told us to get dual 5070 for the same effect.
>>109083124>>109083159>>109083166>>109083200nigga i cant afford that shit
>>109083200every brave user looks like that they probably all tried to scam our grandma and run a 7/11 using tax money
I can barely afford food, rent and Healthcare. What makes you think I can afford this?
>>109083124Mini PCs won.
>>109083124That is 4679 base, but it comes with only 90 days subscription. After that it is another 4.5k each year, or your spark is a paperweight.
>>109083287It's a UMA RAM system, you get 128GB at 273GB/s running on a system that consumes roughly 150W/h. A GPU server is consuming more than that *per* GPU. You're consuming a ton of extra power, creating a bunch of extra heat, and spending more per GB of VRAM with dedicated GPUs. Once you get above 8 or so GPUs the whole cluster becomes very unwieldy and you've basically turned your office into a small datacenter whereas you can just stack 4-8 UMA boxes in a corner with an RDMA switch and consume about as much power as a gaming computer. Buying just one or two sparks/g10s/strix halos is retarded but if you want to run models larger than a few hundred billion parameters it's basically the only viable way.
>>109083124I lack the use cases for "AI desktop supercomputer".
>>109085905this is why nvidia is worth more than india
