CUDA Tutorials I CUDA Compatibility

[@walterppk1989]

I guess this was the full explanation for people who use the CUDA infra directly, but that's not how I usually interact with this stack. As a DS, I've struggled to get CUDA installed on custom docker images in order to develop gpu accelerated tensorflow imgs. The difficulty there lies in the question 'what parts of the CUDA stack do I need, and which version numbers do I need to align, in order to receive GPU support?'. In the end, I defaulted to a tensorflow:tensorflow-gpu image as my base because I couldn't figure it out from a base linux img. I ended up splitting my CI/CD pipeline into a 2 stage build (one where it installs the requirements, and the other to actually add my code), which generates a fairly large final docker image (5gb) and a CI/CD pipeline that takes about 6min when nothing changes. TL;DR: if you need tensorflow gpu support, trying to manually install the required cuda on ubuntu is a pain. Using tensorflow:tensorflow:gpu does the job and frees you from having to worry about any of this underlying cuda architecture stuff, but you do end up with large docker imgs.

[@kevinhamb]

This is exactly what i was looking for! Thanks for the video :)

[@kansairobot2015]

I really need help in Forward compatibility. I have already installed a new UMD on my datacenter. When I do nividia-smi I can see that . I used symbolic links and paths to have the libcuda.so call the correct file. My first question is, this does not include the Cuda Toolkit right? I want to install the Cuda toolkit over there in a separate Conda environment. How do i do that? I tried conda install -c conda-forge cudatoolkit=11.6 and it seemed it went smooth but after that the nvcc still gave me the old version. So my second question is after you set forward compatibility how can I install the cuda toolkit over it?

[@ElinLiu0823]

Hi,beginner asks:Can i learn CUDA just with C basics?Or C and C++ are both required?

[@aswinfamily]

Could i use the application developed in CUDA 10.1 into CUDA 12 or 11.6 ? I'm usually facing kernel driver error..

[@robnertney743]

The driver will work with your application (Backwards Compatibility). If you get some errors, make sure to check to see what they say....there's a chance that you are trying to dynamically link to libraries back in 10.1 that changed in 11.x or 12.x. If you statically link your 10.1 application, it will run on any driver into the future.

[@yonnileung]

The most difficult part to working on Ai stuff is installing the "God will know version combination between Cuda and Tensorflow"

[@yuan.pingchen3056]

could you help me solve the problem of the coda device too expensive ? I'm really want to try the most recent version of CUDA compute capability and want to try write program the tensor core....

[@nneeerrrd]

Two solutions: 1) get a job instead of Fortnite gaming 2) buy lowest level GPU which supports cuda. It will be enough for your first steps

[@yuan.pingchen3056]

​@@nneeerrrd maybe release the software emulated GPU is another solution,just like iphone emulator...., since it can't be called accelerator and should be decelerator.

[@yuan.pingchen3056]

​@@nneeerrrd the most recent CUDA compute capability(Ada Lovelace architecture) did not have product that MSRP below USD150

[@hussienalsafi1149]

❤️❤️❤️❤️❤️❤️❤️❤️❤️❤️❤️❤️

[@xeon789]

i tried everything to make my cuda compatible. i have Nvidia gtx 1660ti (6GB Vram) Notebook Graphics Card. I know for sure it it cuda compatible. [as its mentioned in card spec] i want my use stable diffusion 2.1 web ui with GPU's cuda. please he me figure out correct version for: -------------------------------------------------------------------------------------------------------- nvidia graphics driver: xx cuda driver : xx cudnn driver: xx -------------------------------------------------------------------------------------------------------- please tell which version of above drivers should i install so its works. -------------------------------------------------------------------------------------------------------- when i run below program it returns false, i want it to return true. -------------------------------------------------------------------------------------------------------- import torch torch.cuda.is_available() --------------------------------------------------------------------------------------------------------