A GPU Made of 8,192 Brains and 8,192 Pixels
A homebrew graphics processor built from thousands of tiny RISC-V microcontrollers is a DIY GPU project where each microcontroller acts as its own pixel, forming a massively parallel display and compute array instead of a single monolithic chip. This experiment is less about beating commercial graphics cards and more about stress-testing how far distributed, low-cost hardware can go when thousands of identical processors share the load of rendering simple images in parallel. It turns the usual GPU model inside out, replacing dense silicon with a sprawling field of blinking silicon specks wired into a coordinated swarm.
Electrical engineering YouTuber Bitluni did not set out to build a personal nemesis, yet that is how he now describes his homebrewed GPU cluster. What emerged over six months is a DIY GPU construction fever dream: 8,192 CH570 RISC-V microcontrollers, each about USD 0.13 (approx. RM0.60), soldered onto custom six-layer PCBs and coordinated by another 256 larger cores with FPUs. Instead of relying on one powerful processor, he spreads the work across thousands of RISC-V chips, each soldered to its own RGB LED so that every “core” is also a pixel. This is not a quirky display; it is a statement that parallelism, not brute-force silicon, still has unexplored territory.
Why 2 Million Chips Kill the Full HD Dream
The brutal truth of this homebrew graphics processor is that physics, not imagination, sets the limits. A Full HD panel at 1,920 x 1,080 would need more than two million microcontrollers if every pixel had its own brain. Even at USD 0.13 (approx. RM0.60) per chip, that bill alone would balloon into six figures of components before you even route a single trace. So Bitluni did the rational thing for an irrational project: he scaled down the target to a 320 x 200 QVGA-style resolution that still demands 64,000 microcontrollers when complete.
Today’s prototype is the “small” version: 8,192 chips arranged in 16 x 32 blocks on custom boards. Those boards sit in a circular layout inspired by classic supercomputers, transforming the rig into a dense halo of blinking LEDs. Power alone tells you how far from toy territory this is. Each chip draws about 10 mA, and in aggregate the array pulls roughly 2,161 watts, or 655 amps at 3.3 volts. One quotable fact from this build is that "the current setup pulls around 2,161 watts, or roughly 655 amps at 3.3 volts". If you want a lesson in scaling, this project shouts it: resolution has a price, and it is measured in amps, watts, and sanity.
RISC-V Parallel Processing, One Solder Joint at a Time
Underneath the colorful LED grid is a harshly practical architecture experiment in RISC-V parallel processing. Each CH570 RISC-V MCU runs at 100 MHz with 12 KB of SRAM, and Bitluni wires them in 32 x 32 arrays on blades that push the limits of low-cost PCB fabrication. Groups of 32 of these smaller chips report to a more capable CH32V controller, forming a hierarchy rather than a chaotic flat cluster. That layered design is the only reason this homebrew GPU construction does not collapse into signal noise and timing errors the moment you try to coordinate thousands of nodes.
The journey has been as much about production pain as about parallelism. One early set of blades arrived with several MCUs dead or unreliable, forcing a complete trace redesign to avoid interference. Another failure was classic: crossed MOSI and MISO lines between levels, feeding input into output and vice versa until communication broke. Bitluni jokes that "no matter how hard you try, you will get rx and tx wrong on the first try"—but that is the point. This is what distributed custom chip design looks like when you do not hide behind vendor reference boards: ugly, iterative, and instructive.
From 8,192 to 32,000 Chips: Chasing a 64,000-Core Vision
If 8,192 RISC-V cores sound excessive, Bitluni’s plan makes them look modest. According to his own video, the next version—parts already in hand—is designed for 32,000 MCUs. The full 320 x 200 target calls for 64,000 microcontrollers when complete, meaning today’s cluster is essentially a quarter-scale prototype of a much larger homebrew graphics processor. Everything about the system fights back: board complexity pushed his manufacturer’s online tools to the point that he had to split each blade in two; power delivery required a WS3000 ATX supply plus high-current converters to step 12 volts down to 3.3 volts.
What comes next is not a product launch but a public lab report. Bitluni says he will share a follow-up video detailing what version one can do, along with releasing all design files for anyone “insane enough” to replicate the build. The project remains in its early stages and nowhere near commercial GPUs, but that misses the point. This is graphics as an open experiment: an audacious proof that thousands of tiny, identical RISC-V nodes can cooperate as a GPU, and a reminder that you can still build something unwieldy, borderline absurd, and technically illuminating in a world obsessed with polished, sealed boxes.
