ESP32S3 cluster running 1.58-bit (BitNet) Language model
github.comThis is something I've been fantasizing about for long.
Let's say we took Rust, a language that makes parallelization easier than others (as it helps you avoid some common footguns). How difficult would it be to have a massively parallel computer system made out of many tiny, simple microcontroller-like chips? Let's say we picked many little Risc-V's. Surely this would be an interesting experiment (though I'm not sure whether it'd make economic sense or not...)
It would certainly not make any economic sense, and I guess that's also why noone is seriously looking into stuff like volunteer/enthusiast clusters of home computers to do inference in the same way that e.g. LHC@home works. The main bottleneck for LLMs is still memory bandwidth. Any memory bus not directly soldered on your GPU is terribly slow. That's why one big GPU with twice the VRAM will always perform significantly better than two GPUs with half the VRAM each. And it's also not like you can just solder more memory onto a chip. At modern speeds, the speed of light is a hard limit. For current GDDR7, signals may only travel like 10mm per cycle.
If you spread such a system out over dozens or hundreds of tiny chips, you'll be wasting most of its resources and lose hard to anyone who built a single chip setup.
Mojo might be what you want, especially with "MAX" which is their AI modeling framework, where in other languages you need NVidia's libraries, or AMDs, etc the Max libraries just let you talk directly to the GPU / CPU / ASIC with Mojo. I think Mojo is very underrated in this space right now, but assuming they don't mess it up, it could be a major contender in AI. In theory, if someone releases a board, and Modular (company that 'owns' Mojo) adds it to Max, you're basically in the green to experiment as much as you want.
> How difficult would it be to have a massively parallel computer system made out of many tiny, simple microcontroller-like chips?
It's scaling the communication that becomes hard.
In this project they daisy-chain SPI. I don't believe that would scale very far.
Not exactly what you're describing, and not shipping yet, but a cluster in a box. 8 cores/node, 8 nodes.
See GreenArrays' 144-core Forth chips by Chuck Moore.
Look at Xmos, founded by a transputers-dad. Small uCs, that can be connected together into a massive cluster, while being first-citizen of their xC language.
Man I miss Slashdot's Beowolf culsters
Soon ai in every lightbulb running Kubernetes
Praise the Omnissiah.
With the proliferation of Abominable Intelligence? Quite the contrary!
You know, I've always liked Futurama but I always kind of thought it was silly that literally everything has an AI and a personality.
But, you know, I actually think that there might be a logic to it. Economies of scale might mean that almost-literally every computer you buy in the year 3000 has some kind of AI-assistance chip in there, and sure maybe it will have full AI with a personality spitting out one-liners.
Kind of like how disposable vape pens often have a 24 MHz Cortex-M0+ with 3 kB SRAM and 24 kB flash, which would have seemed ludicrous a while back.
Change the year 3000 to the 2030s and it might be just as accurate.
How many rollingupdate pods does it take to change a lightbulb?
Eventually.
It is a bit of a bummer to see that the degree of 'compression' makes it a fancy llm noise-maker. It is still charming.
haha … this is precisely the kind of project that https://bil-lang.org is aimed at: Go for parallel (ie in this case pipeline processing).
don’t get too excited until we get the TinyGo backend built though ;-)
I'm curious how this would handle grammar checking on a basic word processor. Or maybe generate worlds for small text based games. I have no idea what the capabilities are of a cluster like this.
Seriously though, what are the low cost chips that can usefully run LLMs? Is a Mac Mini the lowest we can go? Are there iGPUs on mini-itx that can do it, or are there dedicated AI chips that one could turn into a pi HAT?
Depends on what you consider to "usefully run LLMs".
Earlier this year, I bought a mini pc from Aliexpress, specs are roughly Ryzen H255, 24GB LPDDR5, 1TB SSD. This was around 350€ including VAT, customs, shipping etc. I would personally consider this somewhat of a lowest class of useful LLM box. It can run 8B models well, up to somewhere around 24B. I currently run Gemma 4 26B A4B Q5 on it, with MTP, and it is quite slow, but smaller models would run okay on it.
Let's teleport back 10 years and what you have is a magic box that could make you billions.
I’m actually working on a small project that’s exactly this! Less quant so it’s only 150M parameters but this is amazing.
Gemma 4 when?
We're gonna need a bigger ESP.
looking forward to try this one https://www.espressif.com/en/products/socs/esp32-s31
"Your scientists were so preoccupied with whether they could, they didn't stop to think if they should"
Thanks for sharing. It's fascinating to see a 0.5B LLM being split across seven ESP32s like this.