Lasse

Running AI at home
2 x DGX Sparks 128GB
1 x RTX PRO 4500 Blackwell 32GB
GitHub:

[Lasse @lasse_ai]

NVIDIA discourages this likely as a way to protect against claims that its usual operating temperatures reduce longevity.

[Lasse @lasse_ai]

If you want to reduce energy usage, increase longevity or want the fans on your DGX Spark to spin up less, a good way to do it is capping the GPU clock to 2,200 MHz.

This was tested using @MiaAI_lab’s single spark Qwen Flash Next recipe.

[Lasse @lasse_ai]

Apply the upper GPU clock limit:
sudo nvidia-smi -lgc 0,2200

Undo it:
sudo nvidia-smi -rgc

[Lasse @lasse_ai]

Probably everyone’s dream in local AI.

I bet this means more great recipes and faster quants.

[Mia @MiaAI_lab]

my dgx sparks are getting reinforcements 👀

[Lasse @lasse_ai]

@devopsfortunato @AgentSparko NVIDIA refused more that this cap would be necessary and emphasised that its usual operating temperatures are within spec.

[Lasse @lasse_ai]

@devopsfortunato @AgentSparko Well people have measured the effect of capping clocks on Sparks and basically at 2200-2300mHz cap the temperature is greatly reduced while the speed drops only by a low digit percentage

[Lasse @lasse_ai]

The king for local video generation is Minimax H3 but depending on your hardware you’re gonna have to run it a little differently.
Honestly letting it create a script that generates the visuals for educational content is also a very good solution, in this case using Manim Community:
https://t.co/I9yGX0hEm6

[Lasse @lasse_ai]

What are the capabilities of local hardware to create educational videos?

I did a little test running @FishAudio's S2-Pro for speech generation on my RTX PRO 4500 and @manimcommunity with some guidelines.

The output far exceeded my expectations and I see great potential in using this as a personal learning resource!

[Lasse @lasse_ai]

I've been following the local AI community for a while now and thought that it’s about time for my obligatory introduction post!

I have spent a lot of time using these two DGX Sparks, even before recent jumps in ease-of-use have been made thanks to primarily @MiaAI_lab

My main concern is still turning these tokens into actual useful work and how to build the right frameworks around them to make that happen; although I recognise a lot of this community has the right “build it and they will come” mindset, because what isn’t useful work is a lot clearer than what is.

Always happy to exchange ideas, configurations, and experiences with others working on local inference.

[Lasse @lasse_ai]

@GrandpaCodes @MiaAI_lab Honestly I’ve got my own gripes about the silly images, but I’m gonna steal Palantirs playbook here and say “Ontology, Ontology, Ontology”

[Lasse @lasse_ai]

@andersjw_ @MiaAI_lab I just got lucky that the founders editions look great on a piano next to a brass lamp

[Lasse @lasse_ai]

@andersjw_ @MiaAI_lab 😆 then you should check the mentions on the last tensorfold release post

[Lasse @lasse_ai]

@sam_sonivalabs @MiaAI_lab True, this configuration was just for the photoshoot 😉

[Lasse @lasse_ai]

@selfhosted_ai @MiaAI_lab nmu

[Lasse @lasse_ai]

@arb5z seems close to double so the price comes out to roughly equal

[Lasse @lasse_ai]

My findings so far have been that GPT-6.1 Terra is actually very proficient at technical work and a whole lot faster than GPT-6.1 Sol.

I would say Terra High is about equivalent to Sol Low

[Mia @MiaAI_lab]

I asked GPT-6.1 Terra to create 100 HTML files

Rules were simple:
- look stunning
- zero repetitive designs
- go full creative mode

I'm going to be short: the results are SAD ☹️

Not only that, GPT 6.1 Terra also failed to follow basic instructions that no other model has failed.

ZERO creativity. ZERO stunning HTMLs.
A LOT of repetitive designs.

One of the WORST results I have ever seen from ANY model. Avoid at all costs in frontend/web design work.

See the results for yourself (incl. prompts):
https://t.co/m4cmva7LL0

All other 100-HTML file experiments:
https://t.co/7V9sIhIBdz

[Lasse @lasse_ai]

My best guess for this design failure is that they optimized this model completely for technical predictable work and even their drafters to enable fast, cheap inference.

[Lasse @lasse_ai]

@zaighia @MiaAI_lab Mia’s GLM-5.3-flash 2 x Spark recipe

[Lasse @lasse_ai]

@TechMDAI I got a PCIe 3.0 riser and 32GB of DDR4 Ram, the only thing I’m gonna run through Strata or Maya are reasoning traces 😂

[Lasse @lasse_ai]

@thomasjuranek @MiaAI_lab There is a lot of people waiting for the Qwen4 release and rightly so.

[Lasse @lasse_ai]

@unclecat19iyb @MiaAI_lab True, it's only a slight bump in single stream decode. It's only decently better for concurrency and higher kv-cache. If you can take advantage of that.

[Lasse @lasse_ai]

@thomasjuranek @MiaAI_lab Agreed, although that is also something a 3090 could do, the question is which tasks you can throw at it that are just that little bit more complicated, that a 27B model couldn't do as easily.