We've moved!
This site is now read-only. You can find our new documentation site and support forum for posting questions here.
Be sure to read our welcome blog!

GATK germlineCNVCaller with GPU and CUDA

Hello I am running GATK germlineCNVCaller on a machine with an Nvidia GPU that seems to be configured correctly as it runs an test program in the GPU. Without the GPU on 16 CPUs it can take 2.5 days to call a CNV on one sample (case mode). I have not been able to get Theano to work with the GPU having tried different versions of Theano, cuda (8, 9 and 10), cudnn etc. One of the error messages suggests that an older version of cudnn is require but I have not been able to get that version to install.

Should I expect the germlineCNVCaller to work with Theano on the GPU or is this something that has not been tested.

Many thanks!

Answers

  • asmirnovasmirnov BroadMember, Broadie, Dev ✭✭

    Hi @davidco We do not provide support for running GATK gCNV on GPUs. It should be possible - however, as you pointed out you need to install older version of cuda, and potentially configure Theano to run on your GPU.

    To cut down on runtime we usually split up input intervals (into blocks of 5k-15k each), and analyze them in parallel. This usually brings the runtime down to 2-3 hours. You can read about how it's done here: https://software.broadinstitute.org/gatk/documentation/article?id=11684

  • davidcodavidco Member

    Thanks. I will try a few more things with the GPU but I am not that hopeful!

    One other thing I have done is to run multiple samples together in case mode in one call the the caller - this is much faster than running them individually.

Sign In or Register to comment.