Four ways to optimize your data center to accommodate AI workloads

AI is expected to transform data centers in many ways, such as changing the data center job market and improving data center monitoring and incident response operations.

However, the biggest impact that artificial intelligence is likely to have on data centers is to change the way data centers work. For those companies that want to make full use of modern artificial intelligence technology, the infrastructure contained in the data center and its management methods must change

The development of AI in the data center will bring a series of developments worth looking forward to Key changes, however specific impact remains to be seen Differences between other types of workloads, such as standard application hosting

While artificial intelligence (AI) workloads come in many forms and have different requirements, most have the following unique requirements:

Requires a lot of computing resources, especially when performing model training.

Benefit from running on bare metal hardware, especially servers with access to GPU resources.

• Resource consumption rates may fluctuate significantly. During the training phase, AI workloads require a large amount of resources, but after training is completed, resource consumption drops significantly in most cases until the model is trained again.
• Ultra-low latency networks are required to make decisions and deliver results in real time.
• Of course, there are other types of workloads that may have these requirements. For example, running artificial intelligence applications and services is not the only use case that can benefit from bare metal servers. But in general, AI software requires more of the above resources than other types of workloads
• Upgrading Data Centers for AI

To optimize facilities for AI workloads, Many data center operators will need to make changes to meet the unique demands of AI. Here are the key data center upgrades in this regard.

Redesign or replace bare metal servers

1. Virtual machines have been the infrastructure resource of choice for hosting workloads over the past decade. However, as demand for bare metal hardware increases for AI applications and services, more and more data center operators may realize the importance of expanding their bare metal offerings. In some ways, this actually is Simplified data center operations. If you run workloads on bare metal, you end up with a less complex hosting stack because you don't have a mix of hypervisors and VM orchestrators.
On the other hand, in order to scale the bare metal infrastructure hosting workloads, the hosting servers and racks in the data center may need to be updated and upgraded. Traditionally, the simplest way to set up servers in the data center was to provision powerful bare metal machines and assign them to any number of virtual machines based on the needs of the workload. However, if workloads need to be run directly on bare metal, more servers may be needed to isolate the workload - meaning the data center will need to replace high-powered servers with smaller ones and update server racks accordingly

Shared GPU-enabled servers

GPU support is not necessarily required for day-to-day operation of AI applications, although GPU servers are used when training AI workloads is beneficial. Therefore, many enterprises only need temporary access to GPU-enabled infrastructure. To meet the needs of enterprises for shared GPU infrastructure, data center operators should consider providing related products. Some enterprises only need GPU-equipped servers in a few cases, so data center operators can temporarily provide access to GPU resources through GPU-as-a-service to better attract enterprises with AI workload needs
Enhanced Network Solutions

Most enterprise data centers already have access to high-performance network infrastructure and provide interconnect services to quickly Data is moved to an external facility. However, to fully realize the power of artificial intelligence, data center network products may need more powerful capabilities. Those enterprises with artificial intelligence workloads need to have two key capabilities: first, they need high-bandwidth network connections, The ability to quickly transfer large amounts of data is especially important when training AI models on distributed infrastructure. Second, the network needs to provide low latency, which is critical for artificial intelligence applications and services that want to achieve real-time execution

3. Higher data center flexibility

Because the resource demands of AI workloads fluctuate significantly, data centers that are more flexible in the amount of infrastructure they support may be needed. AI may also increase demand for services that allow companies to deploy servers on demand in other data centers rather than setting up those servers themselves, because on-demand infrastructure is a good way to account for fluctuations in resource demand.

To this end, data center operators who want to optimize for AI should consider products that make their facilities more flexible. The combination of short-term contracts and services that include more than just rack space where customers can build their own infrastructure may be attractive to organizations that need to deploy AI workloads.

Conclusion

The AI ​​revolution is still unfolding, and it’s too early to know exactly how AI will change the way data centers operate or the type of infrastructure deployed within them. But what is relatively certain is that changes such as GPU-enabled servers and more flexible solutions may become critical in an AI-centric world. Data center operators who want a piece of this pie should make sure to update their facilities to meet the unique requirements of AI workloads.

The above is the detailed content of Four ways to optimize your data center to accommodate AI workloads. For more information, please follow other related articles on the PHP Chinese website!