Humble self-promo here, may I also recommend the team at CentML who dedicated their academic life (PhD and above) to GPU optimizations for high-performance ML/AI to lower the costs.