Most of the database are trained on server with say 16Gb memory where the features fit in memory altogether, since features are at maximum 10Gb for a very large database. So SSD is not really relevant. It is way better to have more cores.
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
Anyone had luck using a SSD drive to speed up acoustic model training, or is most of the time spent crunching numbers?
I have not yet looked at the overall percentage time of disk I/O vs raw CPU number crunching during training.
Thought I would ping everyone to see if there has been some progress here?
Thanks.
Most of the database are trained on server with say 16Gb memory where the features fit in memory altogether, since features are at maximum 10Gb for a very large database. So SSD is not really relevant. It is way better to have more cores.
Already have the 16 Gb memory. Running 8 cores now already. I thought that was the case.
Thanks.