M5 Ultra 80Core GLM-5.3-Flash on DwarfStar Speeds

A user tested GLM-5.3-Flash on an M5 Ultra 256GB 80-core Mac Studio, performing multiple rounds of agentic inferencing. While satisfied with the performance and the large amount of RAM, they noted that the GPU appeared to be a bottleneck. This led them to question the utility of a 512GB unit for AI inference, suggesting the GPU would become an even more significant limitation.

This report uniquely highlights the GPU as a bottleneck for GLM-5.3-Flash on the M5 Ultra, unlike other benchmarks focusing solely on RAM or core count.