4" touchscreen monitoring dgx spark

I bought this 4" screen with esp32 on aliexpress awhile back and decided to make a TUI style dashboard to monitor my dgx cluster.
Inspired by btop++, nv-monitor, and engine metrics by @0rand . Firmware is built with PlatformIO on the Arduino-ESP32 core.

This is what it looks like:

demo

You can find the firmware code on my github repo. There is also a SKILL.md file where you can point your agent to it to compile it for you.

I’m currently designing a 3D printed dock for it. I’ll add it to the repo when I’m done.

Have fun!

Neat, I’m toying with the idea of getting a tablet to display my Grafana dashboard.

So cute!

I might do something like this as well since I have an old amazon fire tablet gathering dust, thanks for the idea!

I have a Pimoroni presto, which has a RP2350 controller. I have not set it up as a monitor app (yet), but I have got GLM5.3-Flash to write code to run on it.

Is yours 240x240 pixels, or 480x480 pixels. I had trouble trying to use 480x480 res on the Presto.

Mine is a 480x480 screen. I ran into hardware quirks in my first try but what really helped was I used deepseek v4.1 flash 1017 (local) to do research and write the spec and had qwen3.8 flash next implement. But Qwen but ran into another hardware quirk which it couldn’t solve and I handed it off to Claude to finish it off.

Seems like the Presto uses the same IPS screen / controller ST7701 as this. I wonder if pointing your agent to my repo as reference to see if it will help you since the RP2350 is also dual core. My repo contains the spec, documented quirks and a devlog. Check out system-d-monitor/docs/hardware-bringup.md at main · dicksondickson/system-d-monitor · GitHub

I skipped the whole LVGL library and opted to use plain platformio + ardruino_eps32 core + C++. Qwen + Claude wrote a custom rasterizer and render which was a much cleaner implementation since I’m just rendering text.

I may well give that a try, I’m thinking my 480x480 lockups are a micropython problem. I may try getting GLM to look at your repo, and see if we can port yours across.

If the issue is Micropython, then the underlying root cause may be memory (it locks up trying to init full-res), so trying a C++ solution may be more memory efficient and more importantly, work.

Thanks for the pointers! Might be a weekend project.

Oh yea I wanted to use micropython too but my agent’s research indicates the screen controller and touch controller is poorly supported in micropython so I ditched that went back to trusty platformio + arduinocore. I used to play with MCUs a lot in the past and its so nice to play with it using AI now.

GLM5.3Flash is just ridiculous. I just gave it some links to presto github and yours, made it come up with a plan, switch to build mode and we have a prototype

Just demo data, it is working on wifi access right now.

Thats awesome! So much fun.

BTW I also have 2x MSI sparks paired with a NUC.

I’m actually going to give your GLM 5.3 flash repo a spin today. Thanks for keep it up to date.

I am still running the v0.6.6 tensorfold python based as my main, but the zig line is slowly (in AI development timeframes) becoming something to prefer. It is just rebasing to v1.2.2 now.

Great idea, I’m gonna set something up like this myself

@giles8 I’m getting great results from your GLM stack. Very fast and great quality. I updated my repo so now it gets proper metrics from your stack.

Great, when I get some time I will get opencode to look at the presto project again.

It’s busy on the GLM testing repo right now, digesting new work elsewhere, and if v.1.2.3 of tensorfold lands, then it will look at that too.

I think the Presto update will be a weekend thing after all

Yeah - all good, glm can take it from here

When I was hand coding micropython with the RP uCs I would always try to offload as much as I could for inputs and instant feedback items to the PIO state machines to keep CPU load down. I wouldn’t have attempted doing a custom PIO interface to a touchscreen but I’m sure GLM can handle that. My devices were always pretty simple with only a few buttons, but I’d have PIO loops running bitmask identification (ripped the idea from old NES controllers) and was able to achieve fantastic UX.

Unfortunately, I am like a really spoilt little brat of a child, who has a model where I can say, I have this:


But I want this:

And incredibly, the moel makes it so… I wrote some very small amounts of micropython manually when I got the presto, it was fun, but I never really got into it, now I am just ridiculously spoilt - I actually look for things online that the model might be able to interact with and build some cool stuff on.

This is the presto version:

@giles8 looks awsome on the presto!

@jrsphd Seems like the recent crop of models are really good at writing code for MCUs. When I was having the LLMs do research to write the spec, I ran Deepseek v4.1 flash, Qwen3.8 Flash Next, and GLM 5.3 Flash, they all seem to come up with similar plans on how to tackle the hardware. It was fascinating to watch.