6

I've been wanting to try out Tensorflow but I don't know if my system has enough hardware requirements for it.

    My system specifications are
  • Intel i3 processor
  • 4 GB RAM
  • 1TB hard disk
  • Nvidia 210 GPU

Is this configuration enough to run a simple AI project like image recognition? I've searched the website and other resources but no where is there any details about system requirements.

bsd
  • 133
  • 1
  • 2
  • 6
  • 2
    Strictly speaking GPU is not required to run tensorflow models. All the computations can be performed on CPU. The real limit to what you can and cannot do using Tensorflow is the amount of memory. If you are going to use GPU it is the amount of GPU memory and if you are going to use CPU it is the amount of RAM you have. As long as your graph, with all of its variables and constants and your data can fit into memory you should be fine. You should be able to do some simple image recognition on your system, like maybe MNIST. Otherwise it all depends on your particular models. – Mad Wombat May 15 '17 at 17:33
  • The CUDA version might be a bit outdated for some of the newer tools but in general you should be able to run TensorFlow although slow. – Maximilian Peters May 15 '17 at 18:50
  • @MadWombat Would there be any issues with running TensorFlow from 32-bit Python? – Kubie Mar 04 '19 at 21:31
  • @MaximilianPeters, just looking for input on above question – Kubie Mar 04 '19 at 21:31
  • the only requirements is the processor need to have AVX instruction, in Intel processors are before sandy bridge. – tiagohbalves May 22 '19 at 17:43
  • I have today come across CPU limitations with TensorFlow. I installed latest TF with pip3, had errors loading libs. Realised TF has problems with Python 3.7 and 32 bit Python. Downgraded Python to 3.6 (64), STILL problems. This thread on SO explains that older CPUs (mine is AMD Phenom apparently without AVX2) need an older version of TF. I ended up installing V1.7 TensorFlow. Then I had problems with my keras_ocr lib. I guess older versions of TF do not support Keras? https://stackoverflow.com/questions/47068709/your-cpu-supports-instructions-that-this-tensorflow-binary-was-not-compiled-to-u#c – Rupert Apr 14 '20 at 21:01
  • I have pentium e2200 processor and xubuntu 18.04 LTS installed as operating system. When i install tensorflow and try to import in a virtual environment, i get an error something called "kernel dumped". Looks like it is because of AVX instruction. – Hilmi Jun 08 '20 at 15:47

1 Answers1

4

UPDATE 2020-01-17 : Try out free online GPU instances. For example colab.research.google.com is one such site and kaggle.com is another! No GPU or local install needed!

I agree with @mad-wombat, there are no minimum requirements to use a CPU for tensorflow.

You said that you are interested in image recognition. With your hardware you could run the MNIST for beginners or the advanced MNIST examples.

Try it out! Tensorflow.org even suggests starting with CPU-only when you first try.

When you become more ambitious and want to do larger projects you'll need to take into consideration how much data you are trying to put into memory. If you are using a CPU or GPU then you'll care about RAM or GPU Memory (respectively), but it certainly isn't a roadblock to starting. Additionally, many people just tune the amount of data stored in memory by doing some of the following :

  • store fewer training examples in memory at one time. Many online examples (like MNIST) start by loading all images into memory first, then iterating on them. This just isn't possible with very large data sets no matter the hardware size. The solution is to stream the data through the network.
  • use smaller minibatch sizes. Of course I recommend using minibatch sizes of 10 to 100, but depending on your image size and number of hyper parameters, you may have to tune this to something lower, but you'll only find that through some trial and error. In the beginning, I wouldn't worry about it.

But you don't have to worry about this when you are starting! - Just install Tensorflow, walk through the tutorials, and get a feel for it. All the tutorials work just fine on GPU!

I do most of my initial testing and coding on a CPU only machine with limited RAM and it works just great.

Good Luck!

Anton Codes
  • 3,663
  • 1
  • 19
  • 28