วีดีโอ

Isn’t the call at 16:27 “call saxpykernel <<<grid,block>>>…” not call saxpy?

A very important issue that MIG introduces to the multi-GPU systems, is the fact that all types of peer-to-peer accesses would be disabled and the communication time across inter-node GPUs would drop dramatically. This will result in an environment in which CUDA-Aware MPI and UCX cannot utilize NVLinks. How will you guys handle this issue? Would users be able to enable MIG on demand? (it usually requires sudo access) On the other hand, enabling MIG statically on even a small subset of the GPUs would sacrifice more performance from the cluster's potential peak, because all of the inter-GPU NVLinks would be useless until MIG is disabled again. How is this going to work?

15:21 MPS and Hyper-Q are separate, but related concepts. Hyper-Q is basically a hardware feature that is available on GPUs since compute capability 3.5. While MPS is a software solution that utilizes this hardware feature to enable multiple processes sharing the GPU more efficiently.

5736 Stark Trace

Thank you so much! I know this is pretty straightforward stuff, but 54:33 really helped me out fix an issue with containerized experiments.

DBInterface is not an interface... I was wondering how you might create an interface in javascript since it's not natively supported. Typescript would be a better choice for this if you really wanted this, but as this is, this "interface" contains implementation, something an interface is not, by definition, supposed to do.

Where's the github you mention would be available?

how can i install these GPU Utilization packages(GPU dashboard)?

Obrigada pelo vídeo incrível! 🇧🇷

Nice explanation. I converted my 2D parallelized fortran code to 3D and now it gives me NaN, any suggestions please?

thanks. One of the better videos on collaborative groups.👌

Be like hay.

Oh, my goodness.

😮 It must not be the right way to do it, because cythin should be at least 40 times faster. He should probably get out of the juoyter notebook and get better, more reasonable measurements.

Is there cuda fortran for windows?

This was nice. Thank you.

hi guys, in 57:11, after "ssh -A -J narval nc10201", do you know why I still got required the password, and then got this error: "Received disconnect from UNKNOWN port 65535:2: Too many authentication failures Disconnected from UNKNOWN port 65535" ? Thank you in advance

Is it possible to automatically get compute node id and connect to it?

Not necessarily new … we did this with the mainframe and the Cray Vector processors … GPUs are great at vector math and solving simultaneous equations

Excellent talk - great content, very informative, very useful! Thank you.

Fantastic presentation, thank you so much! What about xperf on windows environment? Still relevant?

Amazing job!

Thanks for sharing useful contents. It would have been better if you had named it "faculty member edition", cause it is specifically for supervisors.

This deserves 10K thumbs up, not just 39!!!

Nice presentation and useful instructions. Thanks a lot!

Your cython explanation and tutorial is the most crystal clear I've heard...and I have listened to > 10 videos! Excellent content! Thank you for this, it makes things less painful ❤

In the code mentioned at 22:50, the for loop initialisation condition must be g.size()/2 right?. To accomodate any thread group size.

Thank you for the presentation 😊

thanks for the video. Its really helpful I had a doubt, so say you want to train an xgboost model on top of a 200GB dataset and I have a VM with some gpus having a total combined memory of ~100GB (GPU memory + VM memory). Will I be able to train the model successfully on that using LocalCUDACluster .

Great tutorial. May I ask you a question? If I have to use an old version of Node for my project that does not support async/await but supports promises, how can I rewrite the "for" loop on 19:16? I need to sustain the ordering of the promises, which means that just start them in loop with "then()" attached to every one of them is not a solution. Also want to ask you - in one of my old web apps I used an inline <script> block in into which I uploaded a file or some web form for new item creation, that script was something like "if (window.parent) { window.parent.doStuff(data) }". It worked fine even in IE6+, but is it a clean approach?

Great video, helped connect to UMIACS HPC

Thanks a lot. It is a very well structured nice video. Without examples it is very hard to learn programming. Thanks a lot. 😄

It was really a nice demonstration for CUDA. Thanks

Very useful !! specially the explanation of the color coding

is the source code available for the debugger part?

Any idea of when H100 will be available on Graham?

how to profile in production using scalene pls

Do you think the AI programmer community will support/adopt RocM and to an extent AMD DataCenter GPUs?

Can I use this code?

Thank you for the great explanation 🙏

Well done. I hope you'll make a follow up about a larger python program, with classes, various files, etc. The challenge I find is that a profiler is not so helpful once the work spreads out across components. I need something more specific to say "how can this function call be more efficient"?

24:38 How are you configuring Jupyterhub like this in the background connecting to these cluster resources. Is there documentation on this available or open source github that I can take a look at?

awesome. thanks man

'/^seq2$;/

awk '{print $1, $3, $5} filename

Thanks a lot for sharing that. It helped me connect UGent's HPC

How to overcome with this issue:-RuntimeError: NCCL error in: /opt/conda/conda-bld/pytorch_1670525552843/work/torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:1269, i nternal error, NCCL version 2.14.3 ncclInternalError: Internal check failed. Last error: Duplicate GPU detected : rank 0 and rank 1 both on CUDA device 1a000

Thanks for the presentation. Please make MIG available for everyone

Dear Ge, thanks for the video. Dou you know if its possible to debug in many nodes ?

The first example shown will not use multiple GPUs, only one. You can see at 27:36 that in the code (it would have been nice to have line numbers) [device = 'cuda:0'...] meaning it only uses the first GPU. I believe this should be [device ='cuda:0,1'...]