I regret building this $3000 Pi AI cluster
A detailed analysis of building a $3000 Raspberry Pi cluster using 10 Compute Blades with CM5 modules, comparing its performance against more expensive alternatives. The cluster achieved 325 Gflops in HPC benchmarks but struggled with AI workloads, delivering only 0.28 tokens per second for large language models. Despite being energy efficient and compact, the Pi cluster offers poor price-to-performance ratio compared to traditional desktop hardware, making it suitable only for niche use cases requiring high node density or physical separation.