Fifteen Years of FP64 Segmentation, and Why the Blackwell Ultra Breaks the Pattern
Summary
Article tracks the historical FP64:FP32 ratios on Nvidia consumer vs enterprise GPUs from Fermi (2010) to Blackwell Ultra (2026), highlighting deliberate market segmentation that suppressed FP64 on consumer cards. It discusses FP64 emulation approaches (Dekker, Thall), the Ozaki scheme, and how AI workloads are shifting toward low-precision and tensor-core compute, culminating in the Blackwell Ultra reducing FP64 to 1:64 while expanding FP16/FP32 pathways. It suggests segmentation may migrate to low-precision compute in enterprise GPUs and questions future FP64 viability.