The CUDA Handbook ▸ Chapter 11. Streaming Workloads

11.5 Performance and Summary

This chapter covered four different implementations of SAXPY, emphasizing different strategies of data movement:

• Synchronous memcpy to and from device memory,

• Asynchronous memcpy to and from device memory,

• Asynchronous memcpy using streams, and

• Mapped pinned memory.

Table 11-1 summarizes the relative performance of these implementations for 128M floats on a GeForce RTX 3060. Each successive strategy improves on the last, and—despite issuing no asynchronous memcpy calls of its own—mapped pinned memory delivers the highest effective bandwidth.

Table 11-1. SAXPY streaming performance (128M floats, GeForce RTX 3060)