k-Wave

1. luni.jimmy
   

   Member
   Joined: Mar '20
   Posts: 3

   Hi, all,

   When running simulations on the GPU using the MATLAB parallel computing toolbox ("gpuArray"), the maximal number of grid points is limited by GPU memory.

   Since there are 2 RTX 2080 Ti GPUs are available in our lab and they support NVLink, we're considering purchasing a compatible bridge to connect them.

   However, whether or not they can be used as a single GPU (with nearly doubled resources) also depends on software support.

   MATLAB's documentation suggests that its gpuArray variables supports NVLink:
   https://www.mathworks.com/help/parallel-computing/gop.html
   "If x is a gpuArray, data transfer between multiple GPUs in a parallel pool uses fast peer-to-peer communication, including NVLink, if available."
   But we're not sure if the whole simulation process will be supported.

   Has anyone tried this before? Thanks in advance!

   Posted 2 years ago #

2. Bradley Treeby
   

   Developer
   Joined: Mar '10
   Posts: 1,640

   Hi luni,

   In general, I'd advise using kspaceFirstOrder3DG to run the native CUDA code - it will be faster and more memory efficient than using the MATLAB code with the gpuArray option.

   Coming back to your question, I'm not sure - I don't have a multi-GPU system to hand that has NVLink. If you do try, please post back and let us know!

   Brad.

   Posted 2 years ago #

3. luni.jimmy
   

   Member
   Joined: Mar '20
   Posts: 3

   Hi Brad,

   Thanks for your reply and the valuable advice!
   We haven't decided to use NVLink yet.
   If anyone tried please share the experience, thanks again!

   Jimmy

   Posted 2 years ago #

RSS feed for this topic

Reply

You must log in to post.