Very impressive. I recently have a CUDA implementation of HMM. Though I haven't finished all the code. My current observation is that there is about 100x - 1000x speedup. For some ad hoc random data, the CPU code runs over 10 min while GPU code runs within 1 sec.
Update: the project is finished. The code and detailed report can be found at http://code.google.com/p/chmm.
- Written on Fri Apr 17 21:31:43 2009.