linear scan or associative scan

[encounter1997]

Hi, thanks for the clean implementation of the Griffin model!

I have some questions about the scan operation in the Hawk model. The current implementation uses accelerated_scan which performs the associative scan.

However, as discussed in Sec. D in the Appendix of the Griffin paper, linear scan runs much faster than associative scan.

I wonder do you have any plans to implement the Hawk model using linear scan?

[proger]

Hi Wen! I believe this claim is specific to the TPU. Preliminary benchmarks on the GPU have shown that a for loop is slower than a parallel associative scan.