How to Install the Latest TensorFlow Keras with a GPU without CONDA

[@vr4782]

Thank you soooo MUCH, I spent so many hours to figure out why the GPU tf version is not working on my desktop after trying many many tutorials. Yours is the best ever !

[@amiraliCEE]

Thank you very much! Here is a list of the specs that I used for my installation just now (Nov-21) in case it helps others too: I have a 3080 Ti GPU CUDA toolkit 11.5 cuDNN v8.3.0 for CUDA 11.5 And everything else in your instructions worked just great. Thanks for making your courses open and available, and the various installation updates. I've learned so much from your channel!

[@RobertJohnson-xg5kh]

Thanks Jeff! And thanks for keeping these How-To videos up to date!

[@HeatonResearch]

You bet!

[@pioneercolonel]

You can't imagine how much I love you now! Thank you!

[@juanmontenegro7239]

The best professor ever. Wish I could attend your classes. Finally resolved this problem with this video. Bless your soul, professor

[@uly6mlv]

I solved my problem with getting my new 3070 Ti GPU to work. Thank you so much. You saved my day.

[@Fito658]

Thank you very much. I gave up twice yesterday, once today. And this was the video that made tensorflow-gpu work for me! I appreciate it a lot!

[@w.abdalaziz]

Dear sir, what do you mean by (without CONDA) and at the end you great CONDA environment????!) I expected this installation without pre installed anaconda at all.

[@temirlanU]

Omg I met this problem with TF version and totally stuck. And now this is it - a very helpful video)

[@HeatonResearch]

Glad it was helpful.

[@che-weihsieh975]

Thank you for this video! I had so much trouble getting my 3090 recognized. Had to copy one of the cuDNN_64.dll files to the bin directory, it worked out fine eventually.

[@Chuukwudi]

No way! I am not doing this again! This is the reason why I reinstalled Ubuntu. I use a GeForce RTX 2070 Super with Max-Q Design . Installed that Nvidia myself, it worked but my monitor screen occasionally started going blank for a few seconds then comes back on. An attempt to revert back and I messed up with my drivers. Touch pad stopped working, as well as Ethernet, Internet and Bluetooth. I messed up with a lot of things trying to build the tensorflow to work with GPU by building with bazel. Had to reinstall Ubuntu in the end. For now, my motto is, "If old one still works fine, that's all that matters." not all about downloading the latest release.

[@HeatonResearch]

Generally, I agree, however sometimes I need the latest. This video is very much for those that want the cutting edge, conda GPU install is generally safer.

[@Chuukwudi]

@@HeatonResearch Old one has started misbehaving. Something is wrong with my cuDNN. setting GPU growth to True and including some other solutions I saw online are unable to solve mine. I am about to follow this video frame by frame. I hope I do not crash anything.

[@Douken]

Thank you, I got miles ahead with this video. 🙏

[@HeatonResearch]

Glad it helped!

[@Douken]

@@HeatonResearch I wanted to mention that to see the GPU RAM usage on task manager you can scroll down on the performance tab, GPU section. It is below the graphs. And that there’s also a very detailed task manager called Process Explorer that you can download on the Microsoft website. Have an awesome day.

[@mujdatvolkan]

Thank you so much for this informative and helpful video. Without your instructions and advice, I would not be able to install my GPU.

[@Terroid]

Sir, in the title its given download tensorflow *without CONDA*, but at 14:27 you use conda. can i install conda, activate the gpu, and then uninstall conda later?

[@w.abdalaziz]

Misleading video title... :(

[@vincenzodentamaro]

@Jeff how are you doing? I ended up on this video to setup the two A6000 gpus of our laboratory... amazing!

[@HeatonResearch]

Hey Vincenzo, I remember you from the Encog days! Doing good, hope all is well. The A6000 is a great GPU.

[@vincenzodentamaro]

@@HeatonResearch encog is a great framework for Java and its speed compared to several neural networks framework was higher. Learned a ton! Thank you, keep going!

[@hasinthanawod5656]

Thank You! This is the only video found really helpful!

[@shubhamluxkar1348]

Thank You Sir for helping . Love from India👌👌😊😊😊

[@michaelhuppe3293]

Thank you for this amazing tutorial! I have a problem, however. I get the error message: Could not load dynamic library 'cusolver64_10.dll'; dlerror: cusolver64_10.dll not found after some research it says that I for one reinstall everything but with Cuda 11.0 because there is no support for 11.2 for TensorFlow or I should rename the cusolver64_11 file into cusolver64_10 inside but I did that before and it resulted in a memory allocation problem later on. Do you have any idea how to fix that problem? And is it possible to have both 11.0 and 11.2 of Cuda? (or 10.0)

[@benjaminmace31]

Jeff, there is an interesting phenomenon with Conda/Tensorflow and any other installation , the Conda installation for me is sometimes 25-50% faster. You should try a comparison and check. At first I thought it was Linux and Windows then I did Conda install on both and Conda is faster than any other installation.

[@syedalamdar100]

One thanks - My RTX 3090, along with everything, was driving me nuts to get everything to work. This video helps. I would like to know a definitive way to test that my GPU is in use in windows. For some reason, I always see my CPU in high utilization compared to my GPU that only goes to 18% at best.

[@andrewh5640]

You are amazing!! Thanks

[@rohankherath]

Thank you so much ❤

[@neodyo666]

Great ! thats all man

[@shadow1403piros]

Great video Jeff! Thanks for keeping them updated :) I'm building my first DL pc and I've always used AMD cards, so I'm a little confused about all the Nvidia drivers. When you install CUDA the installer says it will overwrite the current display driver so, what is the point in installing nvidia game ready drivers first? I've also read somewhere that CUDA comes with the same drivers (maybe just an older version with better stability). Maybe someone can help me out understanding all this

[@Poossl]

Thank you!!

[@MJ-tn8tw]

Hi Jeff it was working fine till tf version 2.4.1 I accidentally updated my version to 2.5 now it ain't working any fix

[@HeatonResearch]

Hmmm, I will watch out for that.

[@MJ-tn8tw]

@@HeatonResearch hey Jeff thanks for the reply I tried 11.2 11.3 updated visual Studio nvidia graphics display drivers even did -gpu==2.4.1 it is getting dynamic loaded but ain't working getting docker issues but still dynamic library is getting executed my specs is 1060 GTX shall I go for lower versions like cudnn 8.1 and cuda 11.0 ?

[@varun92713]

i tried the same method but at the end i got false. i am using GTX 1060.

[@augustocerberus]

If you have the error "Tensorflow GPU Could not load dynamic library 'cusolver64_10.dll'", uses this solution: stackoverflow.com/questions/65608713/tensorflow-gpu-could-not-load-dynamic-library-cusolver64-10-dll-dlerror-cuso

[@js-gc2hk]

show us the mining for the card

[@HeatonResearch]

Thats coming! Subscribe so you do not miss it. :-)

[@BioDynamicIndustries]

Please do one for Linux!!

[@HeatonResearch]

Sure, I have that here... kzbin.info/www/bejne/mpuQe6Gqr5Jqask

[@zeus7579]

Sorry I yelled at you It turns out that you only need to use CONDA once then you don't need to use CONDA anymore.

[@HeatonResearch]

Oh no problem, but good point, yes.