one gp104 sm has 128 cuda cores and one dispatch units has 32 cores. can i just said that. gp104 sm has 4 dispatch units . and gp100 has 2 dispatch units for it only has 64 cuda cores. Am i right ?
The number of dispatch units per SM is not a fixed ratio to cores across all GPU designs.
In my experience there is typically one dispatch unit per warp scheduler in single-issue designs, and 2 dispatch units per warp scheduler for dual issue designs. The number of warp schedulers can usually be determined from the programming guide. For a direct answer to your question, refer to page 8 here. The corresponding diagram for GP100 is here on page 13. Volta for example (p 13) is a single-issue design.