Is double precision floating point emulated on Kepler for graphics?

Question

Comparing the SMXes on the graphics and compute versions of Kepler, I noticed that there are no units for double precision in the block diagram for the graphics version. The blocks are in yellow on the compute version. Is double precision emulated on the graphics chip? It seems possible, since throughput is only 1/24 of single precision.

Answer 1

NVIDIA has two version of Kepler SM. SM 3.0 targets graphics and non-double precision workloads. The SM only has sufficient data paths for 8 64-bit operations per cycle (x2 if doing DFMA). SM 3.5 (K20) targets high performance compute workloads. SM 3.5 supports 64 64-bit operations per cycle. Both SMs implement double precision operations in hardware.