> Once I had the design [of the enclosure], I tried to 3D print it on my brand new 3D printer, which turned out to be a disaster. Took me some time to learn more about designing for manufacturing, especially for 3D printing.
Would have been nice to hear how specifically it was made to work in the end.
> Turns out, cross-fading two 4MP images at 60 frames per second on a moderately powerful single board computer is not so easy.
Yeah, at first it feels stupid that every pretty LCD screen for bus or train stops, ads or whatnot has a full computer behind it, but then you ballpark the memory bandwidth and realize you need a ~ 1 GHz device anyway just to be able to chew through the pixels fast enough. (1920² px × 24 bpp × 60 Hz = 620 MB/s.) You also realize why “fill rate” used to be such a buzzword 20 years ago and why it took a while until true color became ubiquitous.
On the flip side of the O(n²), you can easily drive a watch-sized display with pretty good ppi using a 32-bit MCU, which is why the Apple/Google/Samsung battery-guzzling approach to smartwatches seems wrongheaded to me compared to Pebble/Zepp/etc.
Large displays are not driven by a general-purpose CPU that bit-bangs a single serial data line, and never were. They're driven by separate circuitry with direct memory access, often integrated on the die of chips meant for these applications. The actual bus to the display controller may be a parallel dot-clock RGB bus, or an ultra-speed differential multi-lane serial (MIPI DSI, HDMI, etc).
And train stop displays certainly don't need 60 fps.
The main constraint for high-resolution displays is memory, not CPU clock speed. Your (odd) 1920x1920x24bpp frame buffer takes up more than 10 MB.
> They're driven by separate circuitry with direct memory access, often integrated on the die of chips meant for these applications
In embedded SoCs the memory bandwidth is often shared between the CPU and the GPU. There is not a separate pool of memory and bandwidth for the GPU.
> The actual bus to the display controller may be a parallel dot-clock RGB bus, or an ultra-speed differential multi-lane serial (MIPI DSI, HDMI, etc).
That’s not what the parent comment is talking about. To get the data into the framebuffer you need to write it first. You also need to read the source images. For blending two images and writing the output, the floor is 3X the total image size every single frame. (Two reads, one write). There are tricks that can be played with frame buffers and scan out blending, but you get the idea.
> And train stop displays certainly don't need 60 fps.
We’re talking about the product in the link, which is trying to blend seamlessly between two images at 60fps.
> The main constraint for high-resolution displays is memory, not CPU clock speed.
I don’t think you understood the problem space.
If you’re doing software blending of 10MB image buffers at 60fps, the CPU can be easily be a bottleneck. You have to ensure the CPU can do enough load operations, math on the loaded data, and store operations to get it done in 16.7ms. Ideally you have a GPU, but if you don’t then you need to pay close attention to optimizations and special instructions to get it done within the frame budget and also allow some cycles for all of the other work the app has to do.
> Your (odd) 1920x1920x24bpp frame buffer takes up more than 10 MB.
The 1920x1920 size came from the article. Parent commenter wasn’t making up odd sizes, they were responding to the article.
Way back in time (1980s) I worked in the broadcast tv industry.
We used to used digital still stores. A single frame of PAL took up 1MB of RAM. The machines had 2MB fitted. The board had a hardware cross fade function implemented in discrete logic chips. They also had a 20 or 40MB scsi drive. It was all controlled by a 6809 cpu. Quite impressive for the time
As a mediatec guy I think the idea that everything needs to be 60 fps is a pest. I understand the idea in gaming, but it is not needed elsewhere. A train stop display can easily do its job at 1 fps or lower. You do not need to display motion at all and the 1 fps you need mostly for blinking things off and on (if there is no dedicated circuit for it).
Most cinema films I watched projected from analog film looked perfect at 24 fps. Heck some anime loocks better at 5 or 10fps than it would look if the motion was fluid at 120 fps. Why would a display at a train stop need double the frame rate of a cinema classic that still looks completely convincing? Because some gamers decided they can't play below 60 fps?
>Most cinema films I watched projected from analog film looked perfect at 24 fps
Until you get a panning shot and then it looks very juddery. Filmmakers have gotten quite good at dealing with the limitation but it is a limitation. (That said, you do have a challenge with higher fidelity video because it makes it harder to hide the imperfections in what the video is capturing, which makes things tend to look more fake)
Anime (and most animation) can adjust the framerate for what makes sense for the shot or even the element. It's common, for example, that panning happens on ones while the animation happens on twos or threes, because low-framerate panning looks bad and it's relatively easy with cell shading to animate the pan at a higher framerate. And again, the animators use a ton of techniques to deal with the low framerate and avoid the situations where it looks bad.
Which could also apply to train stop displays, if the graphics were designed for it (and almost certainly a consistent 24fps will look better than an inconsistent ~50-60fps).
I agree. Even for gaming just the fps is misinformed. It doesn't account for latency and is kind of a arbitrary limit that became industry standard.
Consider a 60fps or 144fps stream can still be out of sync for example you might have experienced watching the same video stream on two different devices and one is more ahead compared to the other while both are the same framerate (one having higher latency from the source).
24fps is good for most real video, animations even dial back to way less animating on 4s, 3s or 2s is common practice. However while gaming you may want higher fps (usually corresponds to lower latency) to allow you to respond in a timely fashion mostly because our brains are quite slow. For embedded applications even if there is movement it should never ever need a 60fps source.
I've finally finished my dream music streamer featuring a vinyl-sleeve-sized square display, a custom carrier PCB for a compute module and a 3D printed case, running a custom-compiled kernel, Alpine mini rootfs and a small C app driving the display.
I did not think that this was doable by a hobbyist at all, let alone using free/open source software only (KiCAD, FreeCAD, VSCode). Turns out I was wrong!
> I never did this before so there was plenty to learn, from basic electronics and magnetics to high-speed signal routing.
How did you learn this?
As a hobbyist I found hardware to be quite impenetrable, given the hard mathematics and physics requirements. My attempts to learn this essentially degenerate to taking an informal engineering course. It quickly becomes clear that I'm years away from making something, and I eventually run out of both free time and executive function. AI's been helping... But I don't have the ability to tell when the output is wrong.
Building something useful in a circuit design using off-the-shelf ICs and components doesn't require a very deep understanding of the theoretical basis for why it works. There's really no need to be gated by thinking that overly complex mathematics and physics concepts must be fully understood before building. Most electronic parts datasheets have example circuits and layouts that can be followed to get a working design. A lot of progress can be made just by reading the datasheet, learning how to use a tool like fusion 360 electronics, kicad, etc., to connect the parts in a schematic, lay them out on a board, generate a gerber, send it to a china PCB shop like PCBway, get the pcb, assemble/solder the parts onto it, some cases may need some firmware which AI can do a lot now, finally fire it up and see if it works. Very few calculations or theoretical knowledge needed, outside of Ohms law.
A purely logic circuit with impedance-matched chips, sure. However, when you need to drive a load, you quickly need to learn about flyback diodes, pullup resistors, and chokes.
You find yourself on forums reading posts like "yeah, that's an n-junction unit, you'll find the circuit much easier with a p-junction. Try ANP1058 and forward-bias the gate", or other such gnomic directives.
"Curiouser and curiouser", said Alice, as she fell down the rabbit-hole.
About a year ago, I found a YouTube channel[0] which consists mostly of a really smart guy (with a proper background in electrical engineering) buying cheap electronics from eBay/Temu/AliExpress/etc, taking them apart and identifying the components, then reverse-engineering them. (Often followed by a discussion of how much better of a product it could be if the manufacturer hadn't cheaped out on a two-cent component.)
I'd previously had a barebones interest in electronics (think: kits with LEDs and buzzers in like 3rd grade), but listening to his discussions and explanations of how these things worked gave me a fairly decent understanding of how some basic electrical principles work.
I also recently watched a video series[1] that explained how microcontrollers (specifically the ATMega328, the one powering the Arduino) work, from a very low level that I've always wondered about. The key part of that series, for me anyway, was getting the chip off of the Arduino PCB and onto a bare breadboard. From there I went and read/skimmed the ATTiny85 datasheet[2] (a very popular "baby brother" chip to the ATMega328) which also helped me understand how these chips do things like PWM on a low level.
I've built one project (a simple PWM dimmer based on a potentiometer) and am working on another (designing a PCB with KiCAD[3] to replace the circuit board for a battery-powered LED lantern, so I can design my own sequencing interface for what happens when you press the singular button), and it has been quite the learning experience, but one I've greatly enjoyed!
[3] This comment is long enough already, but KiCAD is a fantastic piece of FOSS software. It's a full schematic/PCB design suite, and there are plenty of tutorials all across the Internet for it.
> As a hobbyist I found hardware to be quite impenetrable, given the hard mathematics and physics requirements.
IMO there aren't any. It's mostly about reading datasheets, and connecting pins to each other IOC those datasheets.
Maybe the physics and mathematics are more important for analog electronics or designing ASICS?
Do you have any specific projects you'd like to build? Post here and we can talk through it. And/or start by downloading KiCAD and clicking around until you understand the UI.
The programming side is mainly also reading datasheets. Instead of connecting the right pins together, you are writing a certain value to a register, as described in tables.
> It's mostly about reading datasheets, and connecting pins to each other IOC those datasheets .. Maybe the physics and mathematics are more important for analog electronics or designing ASICS?
I agree that electronics is very accessible, and I'll add that in my experience as an EEE the main beginner gotcha is bad PCB layout rather than maths/physics.
For a flight controller say, bad layout would manifest as slightly noisier sensor readings and reduced radio range. But assuming this is 4 layers https://www.anyleaf.org/mercury-g4 then you probably know what you're doing :) - ditto for OP's article.
Simple layout rules will take anyone 99% percent of the way to professional results. Ground planes on inside (or bottom) layers, signals on outside (or top). Don't cut ground planes. Decoupling caps close to ICs. Keep current loops small. Vertical traces on one side, horizontal on the other. A single youtube video could teach you all that.
When I used to do this stuff for a job, math/physics was only needed for analog electronics (interfacing to hardware like specialized/novel sensors), optimizing control systems and signal processing, and general cost/efficiency engineering - but you could go a long way without it.
Just to be clear, I made a working prototype (and put a hyperbole in the title), I have a lot to learn still.
You can learn something practical like "embedded DisplayPort should be impedance controlled to 100 Ohms" without fully understanding the physics behind it.
AI can be very convincing. It told me to put the ESD protection as close the Ethernet connector as possible, so I put it between the port and the magnetics. This worked, but resulted in around 1% packet loss. When I moved it after the magnetics, I got 0% packet loss. This cost me a full revision, but I'd say this is the price to pay for not learning everything by the book.
For GP: this is something you would learn from a PHY manufacturer's EVK schematics and layouts. TI (especially), NXP, Microchip show correct magnetics/ESD part numbers and placement - the trickier thing here is finding an EVK that doesn't just use an integrated MagJack.
I was going to say, AI has really enabled my hardware projects recently. I've found incorrect output is less common with the newest models - and they're capable of drawing decent diagrams too which can really help me sanity check some of their ideas. I've also had some success with asking it to cite sources where possible, so I get lots of useful links into robotics forums like chief delphi and similar which also help sure up the math
I have serious beef with ChatGPT's text-based circuit schematic drawing abilities. Even though the model itself appears to be fully coherent about the big and little picture aspects of whatever's "on bench" I have wasted too much time trying to parse its attempts to draw circuits. I consider them actively harmful in their current form. I quietly hope that one of the LLMs will start actually generating netcode that can be pasted into KiCAD, or even rendering circuit snippets inside of the chat stream properly.
I've seen KiCAD running in the browser, so it's not like this isn't completely doable. It's just a question of resource allocation.
Without any deep knowledge, I was able to build at least a working board with a mcp server [1]. It’s always about the interface. I had to solder a few parts. Yes there where a lot of vias placed. Nice grid but it worked.
Thanks for reminding me that that exists! I'll have to give it a try, as it seems to have matured significantly since I last checked.
That said, outside of casual experimentation, I have almost zero need for an LLM to lay out my circuits for me. It's not just that I do have the deep knowledge, but there's so much more to board layout than routing. What parts I've worked with, what I have in inventory and on my PnP.
I also just really love laying out boards, and I'm in no hurry to give that up.
That said, I've found ChatGPT to be extraordinarily useful for part comparisons, getting second opinions on approaches, figuring out resistor divider values, troubleshooting drivers, and so much more it's difficult to enumerate.
> Turns out, cross-fading two 4MP images at 60 frames per second on a moderately powerful single board computer is not so easy. I needed to use GPU acceleration to make it happen and look smooth.
fizzle fade, munching squares, horizontal/vertical/circular/pinwheel/etc wipes - anything that lets you replace a small number of pixels in the display buffer each frame.
That's incredible! I was thinking about such a project but a quick research didn't show any suitable square-shape LCD monitors. I am glad that turned out to me not true.
I will be happy to participate in the Kickstarter campaign
WDYT about adding something more to the screen than record sleeves, e.g., lyrics display?
Nice design. I have done similar projects for myself and my friends, aesthetics is always the key to a make and forget music system you will actually use daily. My last one re-purposed a pioneer amplifier https://rodyne.com/?p=3380
This is really cool. I would love to switch from using the SoM I'm currently using to the Radxa CM3 but I cannot find even a single unit actually available for purchase on any of the "international" distributor options.
Can you speak to how you're sourcing these, and from where? (And would you sell me one?)
Yes, this was it, looks like it is gone now. According to the CM3 datasheet [0] "Radxa CM3 will remain in production until at least Sep 2029", so there should be inventory hopefully.
A bit tangential, and in the topic of nostalgia for old ways, remember when every block had a photo shop? Couldn’t we get a “PCB shop” in our bright near future? I wonder if the density of enthusiasts that need PCBs made is so low that the entire market is captured by only one or two companies based off China subject to the realities of shipping times.
> I wonder if the density of enthusiasts that need PCBs made is so low that the entire market is captured by only one or two companies based off China
The sad answer is that yes, it is that low.
The better answer is that as PCB manufacturing gets more hands-off, we can look forward to some "PCB totem" somewhere in your city, like those self-service photo printers that exist now.
JCLPCB and who? Having a board made and assembled was both great in speed and effort, and terrible in cost markup when you add tarrifs on a $50 uc at 5 board minimum.
Me too, in the general case, but the square display was a key requirement. And i can see it; op is displaying LP covers; it needs to look good
I think i'd just make the bezel for the display and putty an SBC onto the back with an HDMI converter dongle. I'd likely order a custom picture frame with some rococo moulding, for the anachronistic effect, and display nondescript oils when not streaming.
Easier than it sounds: the app runs a simple http server that serves cover art by track id. The track id is transferred via iOS's built in "Now Playing" metadata fields, which is then used by the device to construct a query and fetch the full resolution image.
Would have been nice to hear how specifically it was made to work in the end.
> Turns out, cross-fading two 4MP images at 60 frames per second on a moderately powerful single board computer is not so easy.
Yeah, at first it feels stupid that every pretty LCD screen for bus or train stops, ads or whatnot has a full computer behind it, but then you ballpark the memory bandwidth and realize you need a ~ 1 GHz device anyway just to be able to chew through the pixels fast enough. (1920² px × 24 bpp × 60 Hz = 620 MB/s.) You also realize why “fill rate” used to be such a buzzword 20 years ago and why it took a while until true color became ubiquitous.
On the flip side of the O(n²), you can easily drive a watch-sized display with pretty good ppi using a 32-bit MCU, which is why the Apple/Google/Samsung battery-guzzling approach to smartwatches seems wrongheaded to me compared to Pebble/Zepp/etc.
reply