optimal hardware configuration to run Kaldi

834 views
Skip to first unread message

Lucian Georgescu

unread,
Oct 29, 2015, 10:45:09 AM10/29/15
to kaldi-help
Hi,

I want to know which hardware configuration is indicated to run Kaldi in an optimal way. I am interested especially to train databases of about 500 hours of speech. Also, I want to use DNN recipes, so I'm interested in GPU computing capacity.

I indicated below the hardware specifications for 2 machines that I use now, but they work very slowly in the training phase.

Please give me a few tips on what it is best to buy having regard to the efficiency / price.
What hardware configuration do you have?

Thanks a lot for your help in the last period!

Lucian


Machine 1:

CPU INFO:

24 processors with these specs:

model name    : Intel(R) Xeon(R) CPU E5-2430 0 @ 2.20GHz
cpu MHz     : 2199.897
cache size    : 15360 KB
cpu cores    : 6


MEM INFO:
MemTotal:       49.41 GB


GPU INFO:
VGA compatible controller [0300]: Matrox Electronics Systems Ltd. G200eR2 [102b:0534] (prog-if 00 [VGA controller])
    Subsystem: Dell Device [1028:04f6]


Machine 2:

8 processors with these specs:

model name    : Intel(R) Xeon(R) CPU           E5630  @ 2.53GHz
cpu MHz        : 1596.000
cache size    : 12288 KB
cpu cores    : 4


MEM INFO:
MemTotal:       49.44 GB


GPU INFO:
VGA compatible controller [0300]: NVIDIA Corporation G92 [GeForce 9800 GT]

Daniel Povey

unread,
Oct 29, 2015, 12:52:08 PM10/29/15
to kaldi-help
I would recommend Tesla K10 or K80 cards, as both have two GPUs per
card. One or two of those cards should be enough to train your models
on with GPUs.
Having plenty of CPU cores is probably a good thing too, as many
phases of training need CPUs.
It might be better to have one big machine e.g. with 16 or 32 cores
and slots for 2 GPU cards, instead of 2 machines, because then you
won't have to worry about things like network throughput.
A reasonably big local disk (e.g. in the range 100-500G) will probably
be a good idea too.
And enough memory-- aim for at least 4G per core.

Dan
> --
> You received this message because you are subscribed to the Google Groups
> "kaldi-help" group.
> To unsubscribe from this group and stop receiving emails from it, send an
> email to kaldi-help+...@googlegroups.com.
> For more options, visit https://groups.google.com/d/optout.
Reply all
Reply to author
Forward
0 new messages