Hardware for Deep learning Explained: Master Your Setup | Course Playlist

Key Takeaways

now it is time to set up your machines and infrastructure for this course which is what we will do in this module but before we do that let us understand the hardware requirements for the computations involved in deep learning and in order to do that let me start with a very simple analogy let us say you want to go from place a which is your home to place B which is your friend's place and you have two options you can either use your Ferrari which happens to be your favorite car or you can use the freight truck now your friend lives only a few miles away so which mode of Transport would you choose in all likelihood you would choose the Ferrari you'll reach to your friend's place fast and efficiently now let's consider another scenario let's say a happens to be your old house and B happens to be your new house and you need to move all of your stuff from house a to house B and you again have the same two options which option would you choose now in all likelihood option b which is a freight truck is a better option because if you choose Ferrari which you can do theoretically you can still choose Ferrari to do this movement but it will end up taking a lot of trips and and obviously take a lot of time and fuel in order to move the stuff which you have a truck is much better option to do this move now why am I discussing Ferrari and truck because this is exactly the same difference between CPUs and gpus so let's Now understand how the choice of Hardware will make a difference in your deep learning algorithms and computations now theoretically speaking you can actually build your deep learning model models on embedded Hardware with low computation Powers as well but this will end up taking months or years of computational time to just build a simple deep learning model and that is where you need specialized Hardware with special compute capabilities so let us try and understand this a bit deeper as mentioned before deep learning is basically complex neural networks the algorithm is nothing but a collection of commands which runs on your processor and out of these commands the most resource intensive task are the metrix multiplication operations a whole series of these operations are performed on input to generate the output these operations are repetitive in nature and have to be performed tens of thousands of time now in order to handle these repetitive Matrix multiplications we need to use the appropriate Hardware and there are three types of Hardware which are available to build these deep learning models namely the central processing units or CPUs which are also called the brain of computers the gpus or graphical processing units which are the backbone of graphical processing or tpus tensor processing units which is newly emerging specialized hardware for mathematical operations so let us take CPUs and gpus and understand the difference between the two now CPUs actually have a few complex scores which focus on doing one task well so these are equivalent of your Ferrari which can go from one place to second place in a very very fast Manner and these are usually used for your general purpose tasks on the other hand gpus have hundreds or thousands of simpler cores which focus on doing Simple tasks in a parallel Manner and these are used in graphical processing or metrix multiplications which can be internally divided into multiple smaller tasks and can be executed in parallel for deep learning we can use gpus to train the models and use their ability to run parallel computations which helps us build deeper models on lot more training data so here is a example of a CPU and a GPU both of which are typically used in medium to high-end laptops or desktops so a CPU would typically have four cores but these are much complex cor so they can handle a large variety of tasks on the other hand a GPU would have close to 3 and 1 12,000 cores and this is not the most advanced GPU the more advanced gpus would offer actually even higher capabilities but the you should note the amount of of course the difference in the two computational hardware and essentially the ability of gpus to run massive parallel operations this is why we end up using gpus for deep learning now tpus are extremely specific Hardware specially designed to run mathematical applications using deep learning they're preferred for deep learning using computational tasks at Google as they are much better at the job than CPUs or gpus we should definitely see a boom in tpus as the technology evolves and matures but right now they cannot be used for general purpose Computing as they do not have as much technological support compared to CPUs and gpus you can still test them on Google cloud system called Google collab which we will see in upcoming video now that you understand the different types of hardware and why they are needed for deep learning we are all set to set up your systems for the course you'll also find a link to a very interesting talk in the resources section which demonstrates the difference between CPU and GPU in a very interesting manner