Kevin

@quinceleaf Enjoy and have fun

#nvidia #spark

@AndeanViber @jun_song The cost-effective switch for up to 8 Sparks is Mikrotik CRS804 DDQ which is $1.2k and the breakout cables can be had for ~$130 each.

@jun_song A guy on YouTube connected 8 DGX Sparks, but the cables are $250 each and switches $4-7k wow. youtu.be/PV89U-PNUNA

Wiring up 4 DGX Sparks is a massive headache and requires an extra $2,000 just for the connectors. I'm going to work on getting GLM5.2 (744b) to run on just 2 nodes instead.

@MiaAI_lab First time local didn't feel like a total compromise

GLM 5.2 is also the best way to use 4x DGX Sparks. It's truly an impressive and a very capable model.

Claude pushing back Fable's removal every week from subs is more about making it feel like unobtainium than anything else. People want what they can't have, and it causes them to place a perceived higher value on it. If it was massively available, it wouldn't be as cool.

@itsankitakar Agentic engineering

Hot take: Vibe coding works better if you already know how to code.

Fable chews through my $200 sub to the point my limits quickly evaporate. Sol does not have this issue, and it's the first time I've used an OpenAI model that feels like it understands frontend. Claude still has a "certain feel" to using it, but happier and happier with Codex.

@hiheyhowareya @NVIDIAAI DGX Spark cluster + RTX Pro 6000s = best home setup

So I heard we’re doing DGX Spark clusters. Have had this guy going since January. Some of the most useful silicon in the building. @NVIDIAAI

claude is good for when you don't know what you want gpt is good for when you know what you want

@Tech2Wild If they made a "50x" $500 plan and "100x" $1000 plan would probably sell really well. Lots of people on multiple $200 plans right now and switching is annoying. If they do it though it probably will be done by reducing further what the baseline "1x" means.

This more than likely wont work. But I am curious if companies like OpenAI and Anthropic will try to force longer plans on to us. Probably not due to competition, but I do know having it where one month you can be a customer and then jump ship because the competitor dropped a model has to bother them.

@MiaAI_lab The only real (optimal?) option besides the Spark for a home lab inference setup is to step up to the RTX Pro 6000s, which unfortunately are incredibly expensive.

@ivanfioravanti @TheAhmadOsman A lot of people ignore the value of privacy. I know some incredibly talented engineers who have refused to even use any cloud provider for inference solely due to not wanting their research and work potentially taken from them.

Now you understand why paying more for Local AI does make sense. @TheAhmadOsman has always been right since the beginning.

Running models locally might not be cost efficient or even as "good" as using frontier cloud subs, but at some point you have to ask yourself what price you put on privacy.

One of my businesses sells a few of our SKUs on Amazon. People routinely return boxes with random garbage in it and Amazon's staff will "inspect it," then mark it "good as new," and let someone else buy it. One SKU we had a 50% return rate, asked for our inventory back, and discovered this.

First time I order from the USA. My ASUS GX10 turn up almost a week early. Great! No. Not great. Seeing a few peoples tweets about issues receiving these from Amazon damaged, I decided to record from grabbing it off of my porch to the unboxing. I suspected bad packaging, possibly damaged, but I didn’t suspect no unit inside. When I initially saw it I noticed some sides of the box the tape was up and the box itself looked like it had been through ww3. I picked it up, it felt a bit light but heavy enough not for my heart to sink too much. Anyway long story short, everyone that knows me knows I keep my packaging pristine. I open the box pull it out, seal was broken, box looked like this and no unit. Charger still inside though guess that’s why it felt a bit heavy. Now amazon asking for a police report and all that bs. Anyone know of anyone with bad experiences like this or how the process works please reach out. What a day.

Zuck: “The pricing from some of the other labs is very extreme and has very high margins. We think that there’s a real ability to be able to offer frontier or very high-level intelligence at a much more affordable cost.” Epic pricing war breaking out among agentic models.

@MiaAI_lab It is nice not having to see this while you are trying to work.

Anyone — no matter how big their account or how important they think they are, who claims the DGX Spark isn’t suitable for real work is either ignorant or a liar. No in-between.

By the way, you’re allowed to build “yet another [whatever]” if it makes you happy. Build whatever you want, who cares.

@TechMDAI Underappreciated how nice it is to be able to leave them on with models loaded while only taking up ~10 watts each.

The DGX Spark AI community on X is absolutely the best. The device allows for rapid expansion of your agent due to high unified memory, 4TB ssd, low maintenance, and low power consumption. Not the fastest, but I don’t have to babysit the device like my main rig.

@johnennis Matches how I use both providers. Claude is great for ideation and brainstorming, casts a wider net. Codex is a more attention to detail-oriented execution tool that won't make assumptions about what you want.

I think where Fable really shines is when somebody has a vague idea what they want to make and they don’t really care what it is Then, yes, it’s going to do something impressive And it is a very smart model But if you know what you want, then I really believe 5.5 is better

@LLMJunky The investment in hardware only gets better with each model release cycle too. Won't ever make sense versus subsidized subscriptions but people doing stuff like this aren't doing it for the cost effectiveness.

I've seen a lot of comments recently about not being able to run models like GLM 5.2 at home. Respectfully, I disagree. But I understand why you feel that way. Because let's be real, its not cheap. 4x RTX6000 Pro is the minimum price of entry if you want a near full precision GLM 5.2 at home at a usable tps. If your argument is well $50K-$60K is an unreasonable amount to spend for AI when you can buy subscriptions instead: then I agree. For most, this is true. But it totally can be done at home.

@filicroval Slippery slope, soon you will want 2, then 4. Enjoy it.

look who just arrived! #DGXSpark #GX10