k-Wave

1. alexalex811
   

   Member
   Joined: Jan '19
   Posts: 2

   Hi,

   My lab is looking at getting a new high performance PC, on which we are likely to run a number of k-wave simulations.

   I was wondering if it is now possible to run simulations across multiple GPU's?

   Alex

   Posted 4 years ago #

2. Jiri Jaros
   

   Developer
   Joined: Feb '12
   Posts: 115

   Hi alexalex811,
   the release of this code is planned for the Q3/Q4 of 2019. We still need to investigate the stability of the new code.

   Best
   Jiri

   Posted 4 years ago #

3. alexalex811
   

   Member
   Joined: Jan '19
   Posts: 2

   Thanks Jiri, thats very helpful.
   Alex.

   Posted 4 years ago #

4. Karthik Nagabhushana
   

   Member
   Joined: Apr '20
   Posts: 6

   Hi,

   I have the same question as Alex. Our team is exploring to use a multi-GPU system such as NVIDIA DGX A100 (8x NVIDIA A100, 40GB memory per GPU) to simulate large domains (1024x1024x1024 and higher). Representing such a domain requires >100GB memory which can only be achieved with multiple GPUs. So, I was wondering if the multi-GPU functionality mentioned in this paper (http://bug.medphys.ucl.ac.uk/papers/2018-Treeby-ISNA.pdf) is supported in the 1.3 version (C++ GPU code)? If not, do you plan to release this functionality in the future?

   In the past, we had used CPU-based compute nodes (128GB memory) to simulate such large domains but want to explore multi-GPU systems to improve our simulation time.

   Thanks

   Karthik

   Posted 2 years ago #

5. Karthik Nagabhushana
   

   Member
   Joined: Apr '20
   Posts: 6

   Our team did get access to two HPC clusters. Cluster 1 is NVIDIA DGX A100 (GPU - 8x NVIDIA A100 (40GB/unit), CPU - Dual AMD Rome 7742 (128 cores @ 2.25GHz)). Cluster 2 is made up of IBM AC922 nodes (GPU - 4x NVIDIA V100 (16GB/unit), CPU - 2x 20-core IBM POWER9 @ 2.4GHz). So, I have two questions - 1) Can I compile the single-GPU kWave source files to create binaries for the above CPU architectures? 2) This is a repeat of the multi-GPU question in the previous comment. We want to make use of the multiple GPUs in a given node to support bigger domains. So, is the multi-GPU source files or binaries available for public use? Thanks!

   Regards,
   Karthik

   Posted 2 years ago #

6. Jiri Jaros
   

   Developer
   Joined: Feb '12
   Posts: 115

   Hi Karthik ,
   1) Yes, you can compile the code for A100. Just install CUDA 11 and add SM architecture 8.0 and 8.6. Check this post
   http://www.k-wave.org/forum/topic/is-nvidia-rtx-3060-compatible-with-k-wave#post-8267
   2) This is still not possible. k-Wave is so much communication bound that the using multiple GPUs don't bring any performance improvement. This is given by the cudaFFT library which doesn't scale.

   Posted 2 years ago #

RSS feed for this topic

Reply

You must log in to post.