PERFORMANCE PER DOLLAR IS GETTING FASTER AND CHEAPER

Summary from the TLDR Dev newsletter

The release of GLM-5.2 shows advancements in inference performance, with 2626 tok/s/node on AMD MI355X at more than twice the cost-efficiency compared to similar NVIDIA setups. Despite historical challenges with AMD's software support, improvements in model optimization are narrowing the gap.