Will DGX Spark do it? Or I must add my hand way deeper into my pocket?

Hi all,

I really need your help on deciding which hardware will be best suited for my needs.

Being an entrepreneur my self and running a small business on my own I usually find myself working literally all day everyday and the problem is that I’m not even productive. The everyday tasks are the ones eating my day and profits eventually…

What I’m looking for, is to automate things on every aspect of my job/projects so I can focus more on me being out there talking to my clients or even have time to think and eventually expand… Rather than organizing files, posting to social, replying to mails or making offers etc…

Now to the point. Use cases that I need the AI to handle.

Secretarial: I need to automate the back office work. This includes tasks like organizing mails, files, folders, replying to mails. Summarizing what was done today and what remains for tomorrow, learn all the brand catalogues so an offer can be also produced etc.

Marketing: Scraping through social, websites etc, finding info and creating campaigns, posting to social, scheduling posts etc. Finding new clients and create land-pages… Video/Image creation is needed on this one.

Automation with Home Assistant: Self explanatory.

Vibe Coding: Able to edit my website/ create new ones. Create apps and anything I may need for my business.

PS: As a side project I vibe code trading algos, so this is a big aspect in which I need for it to perform well.

IT: Currently running two proxmox servers with many vms etc in which I will need this to monitor things and let me know…

Based on the reading I’ve done so far I understand that more or less what I need is N8N automatons: Probably one main AI orchestrator in which I will talk to and confirm every action eg: Through Telegram and some sort of coder like PI or Claude Code and under it several groups of Agents eg: Office, Marketing, Automatons etc with a couple or more agents inside each group.

The problem is that the more I read the more I get overwhelmed and confused and the more I get a feeling that only people which already have some sort of hands on at least and even better work with AI can really point out the correct hardware that will actually do that.

Also realistically the only thing that should respond at least in a normal conversation pace, should be the orchestrator and for the rest they can take their time to develop and finish the project. I mean I don’t necessarily care if the video that will be scheduled for the next social post takes 2 days to be produced…

Ideally since this would run 24/7, I prefer to not go the custom build/2gpus road… especially if I can do more than decently my work with a small form factor/lower wattage option…

Reading this https://dendro-logic.com/engineering/nvidia-dgx-spark-concurrency-benchmark/ I think more and more that Spark may be what I need but really need your opinion and expertise…

Current hardware just in case it can also being combined somehow or get things split…

  1. Laptop with 12th Gen Intel(R) Core™ i7-1280P and 16GB of ram.

  2. Tower: Intel(R) Core™ i7-4790K CPU @ 4.00GHz / 32GB ram / GTX 1070ti (currently converted to proxmox)

  3. Intel(R) Pentium(R) Gold 8505 / 46GB

Is this achievable or am I dreaming?
So for my needs: Multi-modal, Augentic etc, will a DGX Spark/GB10 OEM variant be the best choice? Strix Halo? Any sort of MAC M5? Custom build? Just buy two 3090s and add them to my current build?

Thank you in advance and thank you for taking the time going through all of this.

Your goals are ambitious for sure, but relatively achievable! You have a lot of use cases, and in order to support all of them using only local hardware/models is going to be tough, especially on things like image and video generation. We’re not talking just slow, we’re talking that average “ai slop” you see with messed up fingers and such. Local AI just isn’t there yet for those use cases (to my knowledge).

That said, everything else is on the table. First thing you’ll want to do is just kick the tires on ai in general. Snag a $20 sub for a month with your preferred model provider and see what it can do. You could go Claude Code, but you’ll end up locked into that cli quite quickly. You’re looking for something more like Hermes or Turnstone, full agentic platforms (with more security in mind than OpenClaw) that provide multiple interfaces, all powered by some central coordinator or process. Not sure about Turnstone, but with Hermes you can plug in multiple provider credentials, be it local or cloud, and utilize them in concert.

Any of the spark, strix halo, mac unified memory systems would be a decent starting place if you just want it to do one thing and not really think about much else. Performance is capped in those instances, but it’s your “all in one” solution.

Pick one use case, try it out with local or cloud models, and see if it can do what you actually want it to. Slowly start building out from there

2 Likes

Thank you for your reply.

Going with a sub and generally using a cloud based approach is something I would prefer to skip, mostly because of privacy concerns regarding uploading business/client info etc, but Hermes is something first time I hear and will check it out for sure…

As far as video/photo generation goes, I’m mostly interting on keeping my social engaged by creating creative videos regarding the products I promote, so no people will be “used” for that :stuck_out_tongue: , it will be mostly products taken off catalogues but presented in a better way, so I hope I will not face this “AI Slop” headache.

I did a little bit of research and where I live, a mac M4 max of 48gb is around 3.5k with the 128gb option at 8k :sweat_smile:
a spark costs 6k and the only stirx halo variant is the gmktec evo x2 at 4.6k
So more or less for the AIO option I have to try to find outside my country…

On the other hand I find the AMD r9700 at 1.5k
It got me thinking that if this hardware is enough to run what I need, since I already run 24/7 in proxmox my desktop build i7 4790k, 32gb ram / Z97 gamer pro mobo, I could buy a r9700, setup everything and maybe later add a second one? If this aproach is the best option and could really make me achieve what I need, then I’m ok, otherwise I’m willing to find a way to import an aio system if that would be better comparing to just adding the gpu to my current one.

I can confirm that dual-R9700s is a beastly setup. One of them isn’t quite enough to run Qwen 3.6 35B Q8_0 with the max context (also at Q8_0), you need two for that. However, that still leaves ~24GB VRAM for running another model simultaneously.

Interestingly, Tesla V100s actually have roughly the same TG performance as the R9700, but 30-40% lower PP.

MASSIVELY IMPORTANT NOTE: If you’re going to go with a dual-GPU setup, though, you want a minimum of two PCIE 4.0 x8 slots, not an x16 with an x4 running from the chipset; in the latter case, the R9700 in the x4 slot will see a 30-40% performance penalty. However, you won’t lose any performance by going with an older CPU…I swapped my 13th gen i5 with a 5700X and an X570 board that can do bifurcation (one of the ROG Strix boards), and I get the exact same performance as other folk with R9700s in vastly superior rigs.

With that said, I’m not sure an i7-4790K will be enough. Does it even have enough PCIE lanes?

Well according to Intel
PCI Express Revision: 3.0
PCI Express Configurations: Up to 1x16, 2x8, 1x8+2x4
Max # of PCI Express Lanes: 16

I will wait to gather some more opinions here if any and then probably will go the r9700 route if indeed can offer me what I’m lookin for…
Later I can grap another one!

Thank you for your time and answer!