Your explanation is the right one. If you make Theano use less core, it will slow down, it up on share the core, it will also slowdown as you saw.
There is no fix for this. It is up to you for you want to optimize throughput (total time for many jobs to finish) or latency (time for one job to finish) Fred Le 18 nov. 2016 05:03, "Huijia Wu" <[email protected]> a écrit : > Hi all, > > Recently I found running two or more theano programs on the same GPU will > decrease their speeds significantly (0.5x-0.7x slower for each program). > > I am confused about this. Since my theano program takes about 470MB GPU > memory. The total memory of Tesla K40m > GPU is 11519MB. It would be waste of memory resources when I have to run > one program on one GPU. > > My current guess is the theano program needs a lot of CUDA cores to > compute, and there might be not enough CUDA cores to allocate when running > two > programs. > > Could you explain the reasons and provide solutions? > > Thanks, > Huijia > > > > > -- > > --- > You received this message because you are subscribed to the Google Groups > "theano-users" group. > To unsubscribe from this group and stop receiving emails from it, send an > email to [email protected]. > For more options, visit https://groups.google.com/d/optout. > -- --- You received this message because you are subscribed to the Google Groups "theano-users" group. To unsubscribe from this group and stop receiving emails from it, send an email to [email protected]. For more options, visit https://groups.google.com/d/optout.
