What’s Inside an AI Data Center? Compute, Network, Storage, Power & Cooling

So if you have built a data center which is going to work with AI application or we are which is going to process AI based up data what we need inside it would it be different than traditional data center answer is yes and I would say it is not drastically different but it still has some aspect which is very specific to AIcentric data center. Let's talk about that. So if I go ahead and generalize first that what is inside a data center or let's talk about what is inside an AIcentric data center I would say we would need compute. Compute would ensure that whatever request comes in we can process it or if we have a model needs learning we can create that learning mechanism. So that's why we need compute. Obviously one mod one m one server may not be efficient to have a model which require lot of processing. So what we need now we may need multiple compute nodes and to communicate across them we would need network. So obviously we'll need network in our AIcentric data center so that my nodes can communicate and parallelly work. The data which we have has to be stored somewhere. It could be the existing data or data we are generating net new. So we need storage also in our data center. And obviously these all compute, network and storage cannot function without support infrastructure like we need power, we need cooling, we need security and lot many aspect. So inside this AIcentric data center we have these four fundamental blocks compute nodes, network, storage and support infrastructure. We'll talk about all these as we move forward. Now if you are going to create a data center very specifically for AI, what can be some key constraints when you have to deploy this high density GPU workload. I'm using this term first time here as an explanation. GPU stand for graphical processing unit. We'll talk in detail about how GPUs work and what is inside them. Now when we have these many data center or these many re data to be processed we need huge data center. Obviously that is possible to have but still there would be some constraint which we may have to work with. What we may need what constraints we may have when we are deploying high density GPU workload. First constraint you may come across is power. Obviously there is a limited electricity capacity per rack. In a data center you would have multiple racks. How much power you can provide to one rack? How many racks you can keep into your data center? That will be limitation. So limited electric capacity per rack is important. And when we run this GPUbound workload, they require high consistent power supply. So when you are deploying AI or when you are creating a data center for AI, you need to focus on ensuring that you have sufficient power because it is a key constraint for most data center. Traditional data centers may not be also equipped with great cooling options. So heat generated by dense GPU cluster would be too much. We need to have a better way of cooling capacity of rack and room which can become bottleneck too. So whenever you're designing a data center for AI make sure you have enough electricity. proper cooling arrangement. And third thing obviously you would be needing physical space too. So physical limit may restrict you from expanding. You may need multiple nodes but if there is not physical space available or there is no place to keep your racks then where you would have your infrastructure running. So three things you have to consider when you are deploying AI infrastructure based data center is do you have enough electricity capacity? Do you have proper cooling mechanism and floor space for running your physical server. These are the three key constraint in deploying high density workload.