search

CSTR-Edinburgh / merlin

This is now the official location of the Merlin project.
http://www.cstr.ed.ac.uk/projects/merlin/
Apache License 2.0
1.31k stars 441 forks source link

Hardware Recommendations? #113

Open JRMeyer opened 7 years ago

JRMeyer commented 7 years ago

Dear Merlin Community,

Do you have any recommendations for GPU specifications for training Merlin models?

Also, what would be appropriate for running an already trained model server-side?

Thanks for the help!

-Josh

dreamk73 commented 7 years ago

I guess it depends on your funds. I have run the training on a laptop with a GTX860 and built a desktop with a GTX1080. The second one is obviously much faster. There is some discussion earlier (#78 ) about performance with different GPUs. We don't have a server-side implementation yet so I haven't researched what is needed there. .

m-toman commented 7 years ago

I have to say, for training the regular feedforward DNNs (more or less the SLT demo recipe with a few changes) with about 3k sentences, the CPU does a pretty good jobs. I rented a dedicated server for 30€/month with 2 TB HDD and 24 xeon CPUs and a complete run, including feature extraction and alignment is often done in 5-6 hours. And from that time, a large part are extraction and alignment.

But for large LSTM models a GPU is definitely recommended.