Pete Warden — Practical Applications of TinyML

GPU Hacking

Pete Warden shares his experience hacking the Raspberry Pi's GPU to run neural networks, a task that was both challenging and rewarding. Initially, the Raspberry Pi's GPU was not designed for complex tasks like running shaders, but Pete managed to reverse-engineer it for machine learning applications. He recalls the excitement of getting AlexNet to recognize images in just two seconds, a significant improvement from the initial 30 seconds.

It was some of the most fun I've had in years, because it really was just like trying to string things together with sticky tape and chicken wire.

--- Pete Warden

This project was a testament to the power of creativity and persistence in overcoming hardware limitations. 1 2 3

   

Technical Challenges

Programming on the Raspberry Pi presented numerous technical challenges, especially with its limited GPU capabilities and lack of official support. Pete Warden describes using a text editor to write assembly code, a process that involved manually feeding commands into the assembler and dealing with a minimal instruction set. Debugging was particularly tricky, as he relied on pixel colors to track program progress, despite being colorblind.

Honestly, it was like, in terms of software engineering, it was a disaster, but it worked.

--- Pete Warden

Despite these hurdles, Pete found the process rewarding, often using it as a form of procrastination from other startup challenges. 4 5