Setting Up CUDA, CUDNN, Keras, and TensorFlow on Windows 11 for GPU Deep Learning

[@Virat-Srivastava]

Thankyou Sir... Cant express how happy i am after seeing this work. worked for 3 Days to find a correct guilde for this. 5 windows reset so previously experimented stuff don't mess with new installation. Note that version of Tensor flow compactable with this method is till 2.10.0 . after that you need wsl which whole new level of headache to install. 18:35 at this point just change to install tensorflow==2.10.0 . after this follow everything Jeff says.

[@HeatonResearch]

Thank you for the update, glad it works! This is a really good point, TensorFlow no longer supports native windows install (like this video does) beyond 2.10.0. Future versions need WSL2. I will be posting a video on this next week.

[@DukeOfCurling]

I currently tried everything except for this one...wish me luck...nothing. even that didn't worked. This is tremendously upsetting me that all this process is so freaking hard to get it right. Any other suggestions? PATHs are correct as the tensorflow version too. I installed the latest cuda which is 2.12 nd the cudnn files but no luck

[@cannabaro14612]

​@@HeatonResearcho Please...man.. I need big help to use cuda with 2.11 tf..........T.T

[@steveszeto8968]

@@HeatonResearch Sorry, I can't use it successful and it output the bug of unfound the lib with cudart64_110.dll. However, I have it. Should I install again with it?

[@itsbxntley2970]

@@DukeOfCurling Me too,ended up with false on gpu...think the only option is to try the tensorf version 2.10 or with wsl2 trying that now

[@creative_cooper]

Wow... I spent the whole day yesterday struggling with tensorflow gpu installation. Finally got a solution!. Thanks professor.

[@HeatonResearch]

Great, glad this was helpful.

[@D12075]

19:55 TRUE. Thanks for putting together a straight-forward guide on this. Many Saturdays were ruined in 2020 when I tried to do this, messed up and had start over!

[@NeerajGupta-ot9eq]

Providing simple seamless setups, doing gods work!

[@jamilamuhammad5717]

Jeff, I can't help being excited as this great tutorial set to be my research savior. I have been struggling with this since mid 2021, you are great. Thanks a bunch.

[@tomerbrio4409]

Thank you ! Just a little update regarding the installation of ip kernel. Run the following command: conda install -c anaconda ipykernel

[@abdullahaftab166]

Great! it works, after following the entire video. I used this to install tensorflow correctly and avoid conflicts (pip uninstall tensorflow tensorflow-intel keras tensorboard pip install tensorflow==2.10.0 keras==2.10.0 tensorboard==2.10.0 protobuf==3.19.6 flatbuffers==2.0)

[@kylo404]

it shows there is no 2.10.0

[@abdullahaftab166]

@@kylo404 python version 3.10 not latest one

[@deepanshuchauhan8467]

Ah finally, after 3 days of doing this and doing that i finally found an actual video which helped me to install tensorflow.

[@kylequinn1963]

You are a god-send. I could not for the life of me get CUDA to install properly, turns out I needed to install that MS community edition software first.

[@Stray_Spectre]

I followed along for the cuDNN part trying to get it to work with a different program. I remember banging my head on a wall trying to get it going on my old Win 10 machine and couldn't figure it out on my new Win 11 machine. Adding the environmental variables did it. Your attempt to make this age well worked, as it still works with Cuda 12.1. Thank you so much.

[@meaburro8591]

cannot express how happy i am to have come acros this video. I may have sunk 5+ hours with all the installations. thanks a bunch

[@Andah69]

love how you opened the windows gui to set the paths!

[@sandeepg9921]

Sir, thank you very much for this tutorial. I was performing this task since 6 hours but in vain. At last got to see your tutorial and finally helped me get the GPU for the Project work. Much appreciated, thanks a lot.

[@jams2blues]

You are awesome Jeff, thanks for doing these videos they have helped me significantly

[@driesmichiels4184]

After following this video, if you're kernel stops after you try to train a neural net in jupyter and your anaconda prompt says "Could not locate zlibwapi.dll. Please make sure it is in your library path!" Try to locate this file in “C:\Program Files\Microsoft Office oot\Office16\ODBC Drivers\Salesforce\lib” and copy pasted this file into “C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.6\bin” This worked for me!

[@harshwardhanfartale]

Thank you so much for this video! I followed everything in this video perfectly down to the last-minute detail and it worked for me! Also yes you need to install tensorflow version 2.10.0 for this to work! Thank you!

[@bobparker6769]

Third version of tensorflow installed since 5/20/20, using your instructions; install ran with no issues, many thanks.

[@marc1190]

So for anyone experiencing DLL problems the issue seems to be with the manual install of CUDA (as opposed to conda) which automatically adds some values to PATH. If you select the most recent version of CUDA to install (I chose 12.x) these values are not what TensorFlow 2.10 is expecting hence it can't find the DLLs. Tensorflow 2.10 is required since that's the last version which natively supported GPU. pip install tensorflow==2.10 So if you do want to proceed with the manual install then I'd imagine you need cudatoolkit=11.2 cudnn=8.1.0 in order for it to work. I didn't test this, I just installed them via conda "conda install -c conda-forge cudatoolkit=11.2 cudnn=8.1.0" and its working now. Thanks to @Jeff for helping me see the wood for the trees!

[@patrickhennessy4634]

Thank you so much for this - you made the installation and setup seamless!

[@torfa5309]

I don't advise you guys to follow this guide ...the CUdnn version he installed will not work with the CUDA 11.5 when u are doing an image processing model. you need CUDNN 8.2 with cuda 11.5

[@no-sn3gt]

May I ask what are the full versions you used - for 11.5.x & 8.2.y, what are the x & y? Also what versions of tensorflow and keras did you use? Thanks!

[@torfa5309]

@@no-sn3gt 8.2.0 and the same cuda version he installed in the video

[@torfa5309]

@@no-sn3gt and regarding tensorflow it doesn’t matter which version you are using

[@Ingo-ns3qp]

Sir Jeff, thank you so much! I was able to set up the CUDA stuff (following your detailed explanation) and can run the NVIDIA code samples from GIT. I am really happy! Again, thanks for this video.

[@lingfeizhang9411]

Im getting that the len is 0 (so no GPU found), but with no error message. Would you have any ideas of why that could be the case?

[@alimoghimi8532]

Thanks for your great tutorial video. I had so many problems with setting up tensorflow on GPU before seeing your video.

[@HeatonResearch]

You're welcome!

[@devchoudhary8892]

this is the best tutorial on cuda setup

[@anirudhsrivastava3530]

@ 16:43 PackagesNotFoundError: The following packages are not available from current channels: - python=39

[@surajgupta4438]

It's 3.9 buddy

[@AhmedKhaled-oq9rx]

i can't tell you how much i appreciate what you have done, thank you ♥

[@D12075]

Jeff, it's like you're one or two months ahead of my hardware purchases! I was looking for windows 11 support because I sold my deep learning rig (I overestimated utilization % outside of work, and downshifted to an RTX laptop on sale. For those reading, a dual 3090 setup works quite well if you can find AIO water-cooled cards and acquire them at a good price.) I ran the numbers and for $20-30/month, I can just push my workload to databricks on GCP when I need it.

[@Machinelearning-g8d]

do yourealise the instruction file you share at the end indicates that no GPU available?

[@danielmacedo1910]

Mr. Heaton! Thank you very much!! You are amazing!! It just helped me a lot!! Due to my GPU hardware (GTX family), I had do install precisely as you've shown and worked perfectly!! Oh, guys, I had to use TF 2.10

[@puranjitsingh1782]

I've installed cuda toolkit 12.3 and am not able to run gpu using tensorflow on that

[@Thotokk]

Thank you for the nice explanations, I really spent days trying to fix this! I also got some DLL errors, although I did it exactly as in your video. In another guide they copied all the cudnn files in the Cuda subfolders, which seems cleaner. They also added two other lines into path as v11.7\bin and v11.7\libnvpp and the other two v11.7\extras\CUPTI\lib64 and v11.7\extras\CUPTI\include. That worked for me

[@infiniteworld6919]

Thank you sir you just cleared my concepts due to which i could allocate my gpu in latest version

[@DesXboy]

the \include path at min 12:00 is not the one inside CUPTI or is it?

[@tanhongkiat8196]

Thank you Sir for the detailed tutorials! I have spent much time trouble shooting these issues! Now I am ready to learn more on machine learning!

[@lucysumi6853]

I watched so many tutorials and googled search, finally your video worked for me. Thank you so much.

[@kingtut11]

Jeff... i love you. thanks for providing a thorough af guide.

[@rishabhsheoran6959]

Very helpful video Jeff! Thanks a lot! I was unable to see my GPU for: tf.config.list_physical_devices() I found out that tensorflow does not support gpu after 2.10.0 for Windows so I got that version and python 3.9 or 3.10 and finally everything worked. If anyone is facing the same problem, follow jeff's steps and use the above configuration I mentioned.

[@kylo404]

every thing mentioned is installed properly but the output is coming FALSE also checked line by line dir in cmd but everything is available what is the reason?

[@shubhamojhaa]

i love you with all of my heart. thank you so much for this Jeff!

[@fatemeetsluck]

I have no idea what I'm doing, I've got this Nvidia 4060 RTX Ti, installed the CUDA drivers, and all the python scripts I run always try to allocate on the PC RAM instead of using the GPU.

[@askerycanabarro]

Quick tip. If you have a laptop and your GPU is working fine, please skip step 1, as it can mess up your graphics. For desktops, I do not see a problem.

[@vijayalakshminiranjan276]

Hello, Sir. Thank you for posting the installation video. After installing Visual Studio 2022, I have a question. I installed Cuda 11.5 as you recommended, but it reports that no supported version of Visual Studio was found. Some CUDA toolkit components will not function properly. To get full functionality, you must first install Visual Studio. But I already installed Visual Studio. Could you please assist me with this problem? Reply me ASAP

[@lucuzzo96]

Professor, I think I love you... You saved my day!

[@HeatonResearch]

Glad to hear it!

[@roopanagbhushan7255]

In TF Install Guide , it was mentioned to install MS Visual Studio 2019 , and you have installed 2022 version , won't it cause a problem??

[@roopanagbhushan7255]

Can I just install Visual Studio C++ redistributable only ,, or is the whole MS Visual Studio needs to be installed??

[@saurabhkadam2063]

I have followed all the steps but at the GPU time its getting me as a false what would be the reason

[@bhavinmehta582]

Hello, When i run my model (fit a model) got below error and it stops a kernel. I have followed a same process. Please guide me for the same. 2022󈚪󈚶 14:08:12.877260: I tensorflow/core/platform/cpu_feature_guard.cc:193] This TensorFlow binary is optimized with oneAPI Deep Neural Network Library (oneDNN) to use the following CPU instructions in performance‑critical operations: AVX AVX2 To enable them in other operations, rebuild TensorFlow with the appropriate compiler flags. 2022󈚪󈚶 14:08:13.379986: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1532] Created device /device:GPU:0 with 1661 MB memory: ‑> device: 0, name: NVIDIA GeForce RTX 3050 Ti Laptop GPU, pci bus id: 0000:01:00.0, compute capability: 8.6 2022󈚪󈚶 14:20:58.880634: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1532] Created device /job:localhost/replica:0/task:0/device:GPU:0 with 1661 MB memory: ‑> device: 0, name: NVIDIA GeForce RTX 3050 Ti Laptop GPU, pci bus id: 0000:01:00.0, compute capability: 8.6 2022󈚪󈚶 14:21:46.675136: I tensorflow/stream_executor/cuda/cuda_dnn.cc:384] Loaded cuDNN version 8303

[@ChintanShah-z3w]

hey there! sorry for off the topic question, but which display monitor do you use?

[@timothysimard5767]

Thanks for your attempt to clarify this process. I've been struggling to get this to work for days now. I was hoping that following along your tutorial would do the trick, but no luck. I'm using an NVIDIA GeForce RTX 3060 Laptop GPU on Windows 11. I've experimented with multiple versions of Cuda, cudnn, python, and tensorflow, but nothing has worked. I'm about to throw my computer against a wall. Any chance I could reach out personally? I feel like I need some hand holding at this point.

[@zeuztrumpeter8014]

Experiencing the same problem after watching this, I also have an NVIDIA GeForce RTX 3060 Laptop GPU on Windows 11, any solution yet?

[@timothysimard5767]

@@zeuztrumpeter8014 Unfortunately not. I tried emailing the creator of this video to see if he had suggestions, but got no reply. I've had other things to work on so haven't sunk much more additional time into it. I hate having my own GPU only to use remote computing to run things because I can't figure out how to get this to work.

[@mikeguitar-michelerossi]

@@timothysimard5767 Found any solution? 4060 laptop gpu here.. and it doesn't work

[@jahidtanvir]

Sir, I am getting "kernel appears to have died jupyter notebook" pop-up whenever I try to run model.fit() function. I have installed tensorflow 2.10 following your instruction and official tensorflow pip installation guideline. What should I do? Please Help

[@shmeckles999]

um doesnt tensorflow recommend running wsl 2? that's where I'm stuck on between wanting to use it as OS

[@piyalidas]

i am not able to statrt jupyter note getting following error"book (base) C:\Windows\System32>jupyter notebook Traceback (most recent call last): File "C:\ProgramData\anaconda3\Scripts\jupyter-notebook-script.py", line 6, in from notebook.notebookapp import main ModuleNotFoundError: No module named 'notebook.notebookapp'" how to resolve this , i have googled the error and tried the pip install etc.. stil doesnt work

[@soofishafiya72]

I followed the video step by step, everything working fine until last command resulted as FALSE instead of TRUE. I cross checked the commands in cmd but every path is working. And while opening jupyter tensorflow, it is showing an "error occurred while creating a new notebook". What should be done! Please HELP

[@geleshgomathil3274]

I run into issues of Conda wasnt working with python=39 and not intsaling ipkernal, resolved BY conda config --remove channels conda-forge conda config --add channels conda-forge conda update conda

[@adityay525125]

Although it doesn't help much but torch handles all dependencies via the pip. Just hoping that TF gets to the same kind of support sometime down the line.

[@HeatonResearch]

Yes, agree... installing PyTorch with GPU is a much shorter video.

[@SpiritVector]

I can definitely tell you absolutely love Microsoft buddy.

[@qikeshi1219]

Sorry, can I ask why it shows: "C:\Program Files\Microsoft Visual Studio\2022\Community\VC\Auxiliary\Build\vcvars64.bart 'C:\Program' is not recognized as an internal or external command, operable program or batch file." When I can not run vcvars64?

[@nikemusically]

hi i am having issues with installing the tensor flow

[@NavnilDas-o1n]

Can somebody Please tell is it needed for the VS code C++ Step? I will code using Python only. So what is the need?

[@vinaysagar2321]

I have done everything correct but cannot get success pls someone help me out to connect my gpu to tensorflow, I have checked every enviroment variable one by one every thing is right but still "false"

[@isaackafshkanan509]

Thanks Jeff 🙏! Will Pytorch gpu work aa well or we still need to type " conda install Pytorch torchvision torchaudio cudatoolkit=11.3 cuDNN= 3.2 ?

[@germancruzram]

I have GPU Nvidea 4060TI. What version of cuda would you recommend?

[@eugenesagan5203]

Thank you very much for the tutorial. Does TensorFlow (2.7.0, the latest) really work well with CUDA 11.5? I have CUDA 11.6 installed, but TensorFlow requirements say it needs to be 11.2. So I'm having doubts about whether I should install TensorFlow regardless of that, or find some way to downgrade CUDA.

[@TheShekhar91]

Can you share how you resolved this issue?

[@j3rry837]

So, what's your solution? Did you reinstall windows in the end?

[@dimane7631]

The same issue

[@EmericaTheTour]

Hello, could you do update, how to install everything, because as of today (december 2023), there are versions of TF that don't support GPU. I would love to see, how to get over it right now.

[@AegeanEge35]

Hi what makes difference if I dont use anaconda here? I mean what anaconda ensures easily than other because many people in GPU acceleration considering and giving their examples with anaconda

[@aspboss1973]

I have RTX 3050, every path checks out. Still its showing False... Please help

[@irrelevant_idea]

Thanks very much. Tried many ways but today it worked successfully👍...

[@himanshukandpal5585]

Hey Satyasheel, in the latest tensorflow documentation, i cannot find any paths for cudadnn. did you skipped that part?

[@alexi_space]

how can I run my project from windows with wsl in Visual Studio?

[@sievo666]

Doesnt matter what I do i seem to always get false. (yes i have a nvidia 4090) been through the vid thrre times now

[@mohammadsalehipak3194]

The best video found after one month

[@rabieelhakouni1167]

hi sir ,i have cuda version 12.1 and python 3.10 and tensorflow 2.12.0 when i want to check how many GPU i got False

[@karolstylok542]

What do I need to install for tensor cores and what should I install if I want to use LSTM cells?

[@wildreams]

Do I need visual Studio? Can I use Spyder?

[@yuvisingh12345]

Thank You so much. Finally able to use the GPU

[@DukeOfCurling]

After all this still not working for me I have a 3060 and still return just the cpu. The PATHs are correct, the conda env correctly installed as all modules etc.

[@ik_ben]

Why do we need to install Visual Studio when we use Jupyter at the end ?

[@HeatonResearch]

Good question. Its for the C++ compiler, it is used to compile the CUDA kernels. The actual visual aspects of it are unneeded. However, in windows, I do not believe you can install the VC++ compiler without the IDE.

[@ik_ben]

@@HeatonResearch thanks, it's clear now. At the beginning, i thought you will use the IDE itself.. anyway thanks for this method. It's really the very right way to do things, I spend a week struggling with Tensorflow gpu while trying to install it in Anaconda..

[@ezzal-deenal-hammadi1027]

For the Nvidia Driver Download, mine is GTX 1050 Ti and in options it says either I download Game Ready Drivers or Studio Drivers. My question is which one to install? also I am using Pycharm, so do I need to install jupyter?

[@Apolytus]

I've done everything you do in the video exactly and yet I get a "False".

[@popi3789]

Thank you SO much. I've been looking for a solution for days to force a ML algo to go through my GPU, and I couldn't get tensorflow to detect my GPU accurately. I can tell you the heaven songs played for me all right when I saw the "True" ^^ Also I precise that I had to stick to Cuda 11 (I chose 11.2 according to the compatibility tables of tensorflow/cuda/cudnn) and tensorflow 2.5, because as I understood, tensorflow now doesn't work anymore for windows. Let's see now whether I can launch my script through the GPU in Jupyter...Fingers crossed! Thank you again though :) 🤩

[@gigga5]

The test at the end for GPU still came back False but no error codes about .dll's. I checks all the lines I added in PATH and I could see each one using the 'Dir' command in CMD. How can I switch it to True?

[@Etsoflias]

Did you find the fix for this? I'm running into the same issue

[@cyberversary262]

@@Etsofliasfix ?

[@huanchencai2384]

@@cyberversary262 Try pip install tensorflow==2.10, if you were using tf 2.11, the latest version supports WSL2 but not normal Windows 11 that's what I'm thinking of. Bugged me a whole week. But now works for me.

[@cyberversary262]

@@huanchencai2384 Thanks i tried then same and it wored.

[@lucasbaroneperes1903]

The only video that solved my issue!

[@habibrekik4009]

can someone help guys i am really stuck in just one line when it the anaconda prompt was not entierly shown....

[@Sy06274]

I tensorflow/stream_executor/cuda/cuda_dnn.cc:366] Loaded cuDNN version 8801 Could not load library zlibwapi.dll. Error code 193. Please verify that the library is built correctly for your processor architecture (32-bit, 64-bit) please please please I really need the solution for this please help me out ....

[@surajgupta4438]

Hi Jeff, thank you so much for the video. GPU is been integrated with tensorflow successfully but when running a program kernel died notification is comming and show no module named matplotlib is found and when installing matplotlib my tensorflow GPU is not working. It will be helpful if you can help me on this

[@plumberski8854]

Hope that you can shine some light for me. I just got a 4070 GPU (for Python AI learning). Installed Pycharm, tensorflow, keras, etc. Tested running it with CPU only and with GPU. About 80 times faster with the GPU. What you have shown here is a lot more than what I thought. And the many steps worry me. With the installed Tensorflow, I thought then it will use the CUDA cores. Do I need to do what you have shown to make use of the CUDA? A ball park figure, how many times faster will it be with just tensorflow and with CUDA installed? Many thanks. P/S Thanks for this video. Nvidia's instructions are too advanced for me to understand. Guess due to the many steps, and complexities.

[@md.arrahmandip7638]

Hello Sir, I have a RTX 3060 installed in my pc, I have followed your guide but my gpu is not detecting. Unlike the last part of your video the code returns False. Now Im wondering how to set up my gpu for ML. Thank you

[@VILAB-cs2dq]

Did you solve it..please help

[@timothysimard5767]

I have the same graphics card and have been struggling for days to get this to work. I've tried several tutorials. I must have installed a dozen different versions of cuda, cudnn, visual studio, python, tensorflow, wsl2 etc....all to no avail. Making me crazy.

[@deepanshumusicongo2172]

This video solved a lot of trouble for me .... Thanks

[@sharjeel_mazhar]

Can we uninstall visual studio completely after setting it up?

[@nachohotdawg5634]

It was false for me at first, problem was I had both cuda and cudnn with latest release so I had to downgrade them, then had to unistall tensorflow and reinstall it, and now gpu is true!

[@jiheddachraoui9318]

I have Cuda 12 with cudnn8.6 with the latest driver this doesn't work. can you please help me?

[@LucasOliveira-ek9vu]

Check your TensorFlow version. TensorFlow 2.10 was the last TensorFlow release that supported GPU on native-Windows

[@martinpiceno10]

Thanks a lot, this was very helpful and a great guide.

[@anga6275]

Greetings Sir, i hope you can forgive me for asking but it has been infuriating trying to install cuda alongside drivers in ubuntu 22.04. i would like to see this video for a version of ubuntu.

[@aisolutions834]

print("GPU is", "available" if tf.config.list_physical_devices('GPU') else "NOT AVAILABLE")

[@thatzmey]

Can you pls share the code when you created python tensorflow kernel?

[@productivityacc5128]

If anyone is following this and has a dead kernel when training a model on windows or this msg : Could not load library cudnn_cnn_infer64_8.dll. Error code 126 , you have to install zlib (explained in the nvidia cudnn installation guide)

[@dohanpranataw5543]

hi jeff may i ask something. i cannot found folder c:\tools in my pc, Is this because I haven't installed visual studio yet?

[@dohanpranataw5543]

i had install the visual studio but still cannot find the folder C:\tools

[@kenchang3456]

Similar question as @David White, if you install CUDA, CuDNN on Windows 11 and you want to also use WSL2 do you need to do CUDA, CUDNN installs on WSL2/Ubuntu?

[@HeatonResearch]

See my reply to him. Its on my list to revisit. If you use WSL, you only need the NVIDIA CUDA drive installed at the host level, CUDNN, C++, etc would all be inside of the images. The fact that they are claiming Win11 no longer requires insider edition of Windows means WSL might be an option. Windows random reboots would still be an issue for a serious ML workstation. Generally I unplug internet during a multi-day train in Windows, which is the only really foolproof way I've found to guarantee that the machine won't shut me down and install an update. I could probably refine that approach and just block the Windows update servers.

[@kenchang3456]

@@HeatonResearch DAMN YOU'RE QUICK! And oh, HAPPY NEW YEAR. Best wishes in 2022.

[@amitashnanda3905]

@Jeff Heaton Prof, I tried to follow this process, but I am getting kernel dead error while training my model. Please help me through this.

[@gizem_rick]

did you find a solution?

[@vipinkumaryadav3605]

Thank you, sir. Because of latest version of TensorFlow I struggled a bit. But after using older version it started working. But torch is not able to access GPU any suggestions for torch. Thanks :)

[@DukeOfCurling]

what version of python do you have? mine 2.11 and doesn't work a damn thing. tensorflow is 2.10. I installed the latest cuda which is 2.12 nd the cudnn files but no luck

[@wolfpack4449]

Jeff, I can't see the url for your github install guide. Can you link it here?

[@HeatonResearch]

My bad, just added it to the description. Also here: github.com/jeffheaton/t81_558_deep_learning/blob/master/install/manual_setup2.ipynb

[@wolfpack4449]

@@HeatonResearch Thank you!!