Why GPUs Matter for AI? From Gaming to Machine Learning Explained

— So, compute plays an important aspect. When it comes to compute, what is compute actually? Compute is processing power. Most of us know about compute is CPU. That makes sense, central processing unit that is responsible for performing all the computational thing. But here we are focusing on a AI centric data center. An AI centric data center cannot be imagined without another component of compute which we call GPU, graphic processing unit. What is GPU? We'll talk about that. But first, let's understand why we need GPUs. What is the advantage of having a

GPU? GPUs were first created to ensure that you could render graphics in video game efficiently. So, in animations, 3D environment, images, they can be displayed on a realistic manner and they would be displayed smooth. That started initially from a game called Quake. Not sure how many of you have played it. I have played it. It was first to use 3D accelerator. There were 3D card called 3Dfx and Voodoo which we would buy and install in our PC so that it would perform better for these games. Then, this phenomenon kept on moving forward. That was initial experimentation through Quake and then when it came Unreal Tournament or Quake 3 Arena where it was a deathmatch type of game between player. That is where other aspect of GPUs were used. So, Nvidia that time released the GeForce 256. That was world's first GPU. This is the term coined by Nvi- Nvidia called graphical processing unit. And then, the turn trend continued. Then came Doom in 2004. That was another game and that was where Nvidia enabled or supported something called DirectX shaders which enabled modern graphics realism. So, all these GPU mechanism started to support your gaming need. But then, we realized that GPUs have more potential than just being used for gaming and that's how GPUs have become an integral part of your AI centric data center. So, when did GPU start being used for machine learning? Let's talk about that. Initial process of GPU being utilized, it was

just an idea. So, there was in 2004-2006 Nvidia researchers including Ian Buck who came to Nvidia and then founded all these GPU. They developed Brook, an early idea for using GPU for general purpose computing versus just graphic. So, researchers realized that there is processing power. Why we just needed to use it for graphics? Why can't we use it for additional compute? So, there was a first high-level programming used which was built on as an extension of C that allowed the idea of exploring GPUs. Afterward, based on this idea, in 2006, Nvidia released something called CUDA. C U D A. We'll talk about this in two future sections. It allowed developers to use GPU for general computing. So, once this GPU was there and it was used only by graphic, CUDA allowed it to be used for other general purpose computing and then came breakthrough. People used this into a competition where it's called AlexNet. So, that was a breakthrough. So, that AlexNet trained on GPUs achieved a major leap in image recognition proving GPU power for AI. So, there was a idea that can GPU be utilized? Yes. That got accelerated through Nvidia's CUDA framework and then breakthrough achieved in 2012 where we where breakthrough where AlexNet was trained on GPU and it has won a competition for fastest image recognition. So, that was practical way of implementing things. So, that's how GPU started being utilized for machine learning.