Author Topic: AI programmers GPU question (TensorFlow)  (Read 123 times)

Offline PopeyesPappy

AI programmers GPU question (TensorFlow)
« on: November 17, 2020, 12:11:15 PM »
My brother needs a new computer. He is semi-retire, does AI type programming for fun, his current PC isn't up to the task. We are planning on buying and/or building him something for Christmas, and I need some feedback on GPU's. Apparently what he cares about is TensorFlow which I know absofuckinglutely nothing about.

I think I have 2 choices in about the same price range. GeForce RTX 3070 versus Quadro RTX 4000. The 3070 is built for gaming. The 4000 appears to be purpose built for his application. BUT the 3070 has twice the number of cores as the 4000. Any feedback on which direction I need to go here would be mucho appreciated. I can't ask my brother because this is a surprise.
Save a life. Adopt a Greyhound.


Offline Baruch

Re: AI programmers GPU question (TensorFlow)
« Reply #1 on: November 17, 2020, 12:32:01 PM »
Interesting idea!!  I was just looking into TensorFlow last month ;-)  Machine learning is a fascinating area to explore, and ...

https://www.pugetsystems.com/labs/hpc/How-to-Install-TensorFlow-with-GPU-Support-on-Windows-10-Without-Installing-CUDA-UPDATED-1419/

That is an example of installing TensorFlow on Win10.  Ubuntu, Windows, macOS, and the Raspberry Pi ... all can run TensorFlow (which is used for Machine Learning (ML) projects).  Swift For TensorFlow is a programmable extension to TensorFlow.  While Swift is usually run on macOS, Swift For TensorFlow can be run in Windows too.

"To recap the differences between the two: Machine learning uses algorithms to parse data, learn from that data, and make informed decisions based on what it has learned. ... While both fall under the broad category of artificial intelligence, deep learning is what powers the most human-like artificial intelligence."

AKA check the install doc for TensorFlow or Swift For TensorFlow to determine if your hardware is good enough!!  Always choose the software first, then the hardware ... I am Baruch, complete idiot and I know nothing (like Sgt Schultz).
« Last Edit: November 17, 2020, 12:34:04 PM by Baruch »
Ha’át’íísh baa naniná?
Azee’ ła’ish nanídį́į́h?
Táadoo ánít’iní.
What are you doing?
Are you taking any medications?
Don't do that.

Offline PopeyesPappy

Re: AI programmers GPU question (TensorFlow)
« Reply #2 on: November 17, 2020, 02:59:54 PM »
He's a Linux guy all the way.

Meanwhile someone pointed me at this: https://timdettmers.com/2020/09/07/which-gpu-for-deep-learning/#Do_Not_Buy_These_GPUs

Which contained a direct answer to my question, but may have some other stuff in there you might be interested in.
Save a life. Adopt a Greyhound.


Offline Baruch

Re: AI programmers GPU question (TensorFlow)
« Reply #3 on: November 17, 2020, 05:10:58 PM »
He's a Linux guy all the way.

Meanwhile someone pointed me at this: https://timdettmers.com/2020/09/07/which-gpu-for-deep-learning/#Do_Not_Buy_These_GPUs

Which contained a direct answer to my question, but may have some other stuff in there you might be interested in.

There are multiple flavors of Linux, Ubuntu is probably a good choice.

URL I see is a good recent one.  On highly souped-up PCs I would contact a specialist in some college town where they teach/research this stuff (like this guy).  I can only understand standard software/hardware.  installation of specialized software on specialized hardware would be a bitch in my experience.  If there isn't a wizard (app) that can reliably do it, again I would have a specialist do the app installation.

Good luck on stuffing that stocking!
« Last Edit: November 17, 2020, 05:12:33 PM by Baruch »
Ha’át’íísh baa naniná?
Azee’ ła’ish nanídį́į́h?
Táadoo ánít’iní.
What are you doing?
Are you taking any medications?
Don't do that.

Offline PopeyesPappy

Re: AI programmers GPU question (TensorFlow)
« Reply #4 on: November 17, 2020, 07:37:09 PM »
This is kind of a back to his roots movement for my brother. He concentrated on parallel processing while he was in school in the 80's and early 90's. Then worked on autonomous behaviors in the late 90's and early 00's. He just finished a paper he has been thinking about for years, and now that he finished that up he decided to start getting more into machine learning. All way over my head...

I decided I needed to talk to him about hardware though. So much for a surprise. I need to know if he would rather have a low end $2,000-$2,500 complete workstation he can use now and maybe add a second GPU to later, or start collecting parts for a more expensive 4x GPU workstation later.
Save a life. Adopt a Greyhound.


Offline Baruch

Re: AI programmers GPU question (TensorFlow)
« Reply #5 on: November 17, 2020, 09:56:13 PM »
This is kind of a back to his roots movement for my brother. He concentrated on parallel processing while he was in school in the 80's and early 90's. Then worked on autonomous behaviors in the late 90's and early 00's. He just finished a paper he has been thinking about for years, and now that he finished that up he decided to start getting more into machine learning. All way over my head...

I decided I needed to talk to him about hardware though. So much for a surprise. I need to know if he would rather have a low end $2,000-$2,500 complete workstation he can use now and maybe add a second GPU to later, or start collecting parts for a more expensive 4x GPU workstation later.

Two forms of ML ... deep neural nets and big-data.  You can do deep neural nets at home, but for big-data you have to have access to big data ... like a giant corporation or a government.  Your price range seems reasonable given last year a guy tried to sell me a full custom PC with all the bells and whistles.  I was curious about ML a year ago (and many other things), but didn't care to spend my time on it.
Ha’át’íísh baa naniná?
Azee’ ła’ish nanídį́į́h?
Táadoo ánít’iní.
What are you doing?
Are you taking any medications?
Don't do that.

Offline Baruch

Re: AI programmers GPU question (TensorFlow)
« Reply #6 on: November 26, 2020, 02:28:51 AM »


State of the art ... has moved from TensorFlow to PyTorch
Ha’át’íísh baa naniná?
Azee’ ła’ish nanídį́į́h?
Táadoo ánít’iní.
What are you doing?
Are you taking any medications?
Don't do that.

Offline PopeyesPappy

Re: AI programmers GPU question (TensorFlow)
« Reply #7 on: November 26, 2020, 08:26:03 AM »
Facebook vs Google. Python vs JavaScript and Swift.
Save a life. Adopt a Greyhound.


Offline Baruch

Re: AI programmers GPU question (TensorFlow)
« Reply #8 on: November 26, 2020, 10:16:31 AM »
Facebook vs Google. Python vs JavaScript and Swift.

Happy holidays for you and your brother.  The problem for these giants is even worse than for the companies I worked for, they are probably maintaining a trillion lines of code ;-( ... the cost of porting to a new paradigm would cost more than all the tea in communist China.
Ha’át’íísh baa naniná?
Azee’ ła’ish nanídį́į́h?
Táadoo ánít’iní.
What are you doing?
Are you taking any medications?
Don't do that.

Offline Baruch

Re: AI programmers GPU question (TensorFlow)
« Reply #9 on: November 30, 2020, 01:03:01 AM »
In the automotive application there is preprocessing and postprocessing of data vs the trained AI TensorFlow end result).  This involves data fault vs data processing fault tolerance.  You might want to share this video URL with your brother to provide context ,,,



This is why I have considered learning Agda.  The video also mentions a concurrent fault tolerant version of C++ called Rust.
Ha’át’íísh baa naniná?
Azee’ ła’ish nanídį́į́h?
Táadoo ánít’iní.
What are you doing?
Are you taking any medications?
Don't do that.