A GPU-powered Pi for more efficient AI?

The Raspberry Pi is a compelling low-power option for running GPU-accelerated LLMs locally. For my main test setup, here's the hardware I used (some links are affiliate links): - Raspberry Pi 5 8GB ($80): https://www.raspberrypi.com/products/raspberry-pi-5/ - Raspberry Pi 27W Power Supply ($14): https://www.raspberrypi.com/products/power-supply/ - 1TB USB SSD ($64): https://amzn.to/3OjJysQ - Pineboards HatDrive! Bottom ($20): https://amzn.to/3Zbz0T5 - JMT M.2 Key to PCIe eGPU Dock ($55): https://amzn.to/4eCpi0g - OCuLink cable ($20): https://amzn.to/3YTXNJW - Lian-Li SFX 750W PSU ($130): https://amzn.to/48T4a4R - AMD RX 6700 XT ($400): https://amzn.to/3UXywgI And here are the resources I mentioned for setting up your own GPU-accelerated Pi: - Blog post with AMD GPU setup instructions: https://www.jeffgeerling.com/blog/2024/amd-radeon-pro-w7700-running-on-raspberry-pi - Blog post with llama.cpp Vulkan instructions: https://www.jeffgeerling.com/blog/2024/llms-accelerated-egpu-on-raspberry-pi-5 - Llama Benchmarking issue: https://github.com/geerlingguy/ollama-benchmark/issues/1 - AMD not supporting ROCm on Arm: https://github.com/ROCm/ROCm/issues/3960 - Raspberry Pi PCIe Database: https://pipci.jeffgeerling.com - Home Assistant Voice Control: https://www.home-assistant.io/voice_control/ - James Mackenzie's video with RX 580: https://www.youtube.com/watch?v=J0z09Ddr58w Support me on Patreon: https://www.patreon.com/geerlingguy Sponsor me on GitHub: https://github.com/sponsors/geerlingguy Merch: https://www.redshirtjeff.com 2nd Channel: https://www.youtube.com/@GeerlingEngineering 3rd Channel: https://www.youtube.com/@Level2Jeff Contents: 00:00 - Why do this on a Pi 01:33 - Should I even try? 02:06 - Hardware setup 04:34 - Comparisons with Llama 05:43 - How much is too much? 06:52 - Benchmark results 07:41 - Software setup 09:13 - More models, more testing