SemiAnalysis says AMD’s DeepSeek v4.1 Flash image trails Nvidia on performance per dollar
SemiAnalysis says AMD’s image works out of the box but was released two days after CUDA vLLM support—a gap it frames as evidence of Nvidia’s software advantage.
TLDR
In its September 12, 2026 assessment, SemiAnalysis said AMD’s DeepSeek v4.1 Flash image delivered “up to 14.8x worse” performance per dollar than Nvidia’s H200 and “up to 42x worse” than B200/B300. It said AMD’s release followed CUDA vLLM support by two days. The publisher credited Nvidia’s collaboration with its six-million-developer ecosystem for day-zero optimization, arguing that rapid model support is CUDA’s competitive advantage.
Combined views
111.6K
1 Source, first seen 19d ago