The Changing Role of the Data Center in Design

With the introduction of new compute options for design and simulation workflows—more powerful CPUs and GPUs for the desktop, cloud computing/high-performance computing (HPC) appliances—engineering organizations have more options than ever when deciding how they handle simulation and modeling tasks.

As desktop workstations become more powerful and organizations tap into cloud-based resources for some operations, the role of the on-site data center is evolving. Digital Engineering reached out to a variety of industry experts to find out how engineering organizations are balancing their use of these different compute options, and how that is affecting the utilization of data centers.

Given the various options for compute in design/engineering workflows—HPC, cloud, workstations, etc.—how is the role of the data center changing for engineering organizations?

Rod Mach, TotalCAE: As more and more general-purpose IT workflows are moving to SaaS [software-as-a-service] and cloud-based offerings, we see the focus on data centers to be primarily around high-performance computing [HPC] and artificial intelligence [AI].

In many cases, HPC for CAE [computer-aided engineering] simulation and/or AI is being moved by our clients to “co-location” facilities as data centers are being wound down and no longer available for CAE teams. A co-location facility (or “colo”) is a professionally managed data center where businesses can rent space to house their own servers, storage, and other IT equipment, allowing them to access the data center’s infrastructure like power, cooling, and network connectivity without having to build and maintain their own facility. These colo facilities have fast data access into the cloud vendors, which facilitates hybrid deployments.

Michael McNerney, Vice President of Marketing and Network Security at Supermicro: The data center of the future for manufacturing organizations will need to be designed or modified to incorporate more AI-capable servers. The highest-performing of this new generation of servers will require liquid cooling to dissipate the heat that is generated by the GPUs and CPUs. In addition, these new data centers will have to be able to work seamlessly with edge devices, for instance, on the factory floor to monitor operations and predict failures.

Wim Slagter, Director of Partner Programs, Ansys: The role of the data center in the engineering organizations that we deal with is evolving significantly. This is because compute options have diversified across HPC and cloud and workstations. Data centers traditionally served as a centralized hub for HPC workloads, but the use of cloud computing and workstations is transforming that model. Data centers are now integral to hybrid environments, seamlessly integrating HPC with cloud resources to allow organizations to dynamically scale computing capacity based on workload demands. Giving geographically distributed teams centralized access to compute resources can help them collaborate more effectively and efficiently using very large simulation data sets.

What are some of the pros and cons of maintaining an onsite data center for these companies?

Michael McNerney, Supermicro: Onsite (or on-prem) data centers offers a number of advantages for manufacturing organizations. The ability to size the IT infrastructure to the needs of the entire organization is crucial to reduce the [total cost of ownership] for the company. Policies can be put in place to allocate resources based on the needs at a point in time, which can also reduce overall costs. Specific configurations of systems can be implemented based on application profiles. Also, security policies can be put in place on the on-prem data center to protect sensitive data during the entire life cycle.

Wim Slagter, Ansys: Not every organization is able to maintain a data center, but in our partner ecosystem there are companies that specialize in providing services to our customers if they don’t have the knowledge in-house or expertise to do so.

We provide HPC Platform Services to our customers to make it easier to deploy hybrid cloud environments, combining on-premises hardware with the public cloud. You can get the best of both worlds. With on-prem resources, you have greater control over sensitive data, which is an advantage. Our cloud marketplace solutions on AWS and Microsoft Azure make it easier to set up your own virtual machines on the public cloud that are tailored to very specific applications and optimized for performance and cost effectiveness.

Rod Mach, TotalCAE: Eliminating the cost of maintaining an existing data center for a small number of workloads is the main “con” of maintaining the data center cited by our clients who have eliminated their data centers and moved to different facilities.

The main reasons we see our clients still utilizing a data center or co-location facility for HPC/AI is that it doesn’t always make sense to pay a cloud premium to support workloads that do not need infinite cloud scaling, are not cloud-native, and have large data transfer requirements, which describe most CAE codes today.

Today you can now get 384 cores in a single server, or 8x NVIDIA H200 GPUs in a single server. For many clients, the horsepower of even a single modern server is feasible to own and operate on-premises without needing to worry about data transfer issues and possible capacity constraints, with the new constraint being whether they have enough CAE licenses to utilize the capacity of modern hardware.

Many of our clients are doing hybrid deployments where we deploy in their cloud provider for the flexibility of cloud, while still doing on-prem to gain the data transfer, guaranteed capacity, and cost advantages of on-prem coupled with cloud flexibility. We have seen, with the announced shutdown of some cloud CAE SaaS providers, an interest in having more control of their workloads to avoid disruption through utilizing their own infrastructure either on-prem or in their own cloud.

How is new hardware changing some of those trade-offs? What advancements are improving data center operations?

Rod Mach, TotalCAE: Today modern HPC and AI equipment power requirements mean most clients with very large deployments are power and/or cooling constrained. New liquid cooling technologies help go beyond the 40kW per rack limit of air-cooling technologies. For air-only environments, heat sequestration solutions enable clients to gain back rack densities they may desire.

A major shift caused by new hardware is the sheer power of a single server today dwarfs the requirements for many clients. For example, a single eight-way NVIDIA H200 server can easily handle a 322M-cell CFD model. On the CPU side, today you can get a single server with 384 CPU cores. These advances mean that the amount of physical equipment required to solve even the largest CAE models has been dramatically pared down, such that most any client with an existing data center can accommodate these types of servers on the small- to departmental-cluster sizes. With today’s modern hardware, most clients are CAE license bound, not compute bound.

Michael McNerney, Supermicro: The new hardware, which is capable of AI and HPC workloads, is requiring onsite data centers to plan for liquid cooling. While the work per watt is steadily increasing with each new generation of servers, the workloads are increasing as well. New CPUs have more cores than ever before, which speeds up many applications. While direct-to-chip liquid cooling is sufficient for many data centers today, immersion cooling should be investigated.

Wim Slagter, Ansys: The low power CPUs such as ARM-based processors offer very good performance for Ansys Fluent and LS-DYNA, with up to 60% less energy demand, and these processors are easily accessible on Google Cloud, Azure, and AWS. In my opinion, that makes the cloud option more competitive with on-site HPC.

GPUs have gained popularity for their exceptional processing capabilities, and there are 19 Ansys software products that can be accelerated on NVIDIA GPUs, for example. Where users need to integrate GPU products with third-party, CPU-only apps, leveraging cloud computing gives them the most flexible and efficient solution to accommodate these diverse computational requirements.

Lower power CPUs, GPUs, and liquid cooling advancements are clearly improving data center operations. Cooling and reduced energy consumption allow denser deployments in smaller space, which makes on-site HPC competitive with cloud-based solutions, as well.

In your experience, how are engineering companies deciding what types of compute resources will handle what types of workloads, i.e., workstation vs. data center, or on-prem data center vs. cloud?

Michael McNerney, Supermicro: The trade-offs between different types of computing devices can be complex. Workstations are ideal when working interactively and the local graphics performance can keep up with the demands of the user. However, there are also instances where a remote server can handle the graphics demands for lower performance requirements. A centralized server can be used for lightweight screen tasks and can be shared among a number of users (VDI [virtual desktop infrastructure]). A server or rack of servers in the data center is the best choice for running simulations, due to the high core counts, higher clock rates, more memory, and faster networking than is typically installed in a workstation.

Rod Mach, TotalCAE: If a workstation is getting sufficient turnaround time and there is not a requirement to schedule licensing tokens, workstations can be found for small and quick models. Once you get sufficient engineers and model complexity in the several hours range, it often makes sense to offload those to a HPC cluster or cloud resources. Often the decision comes down to budget, how good internet bandwidth is to their cloud provider, and the frequency of simulation. We have simple calculators to compare the different options to help clients decide the optimal mix, and most clients have both on-prem and cloud and make the choice based on turnaround time requirements.

Wim Slagter, Ansys: Cost is a key factor. Workstations come with a lower cost for a given task, but may lack performance while increasing IT operational costs. With onsite data centers, you can overcome that but they require a higher initial investment and come with predictable, ongoing costs for persistent workloads. Cloud is preferred for fluctuating demands and workloads, but potentially increases your cost over time for persistent users.

Another factor is data security. Simulation is often an IP consideration, and certain simulation workloads remain on-prem because they require stricter control over security protocols. There is also an aspect of collaboration. While workstations are optimal for individual tasks, an onsite data center or cloud-based resources can facilitate collaboration across geographically distributed teams.

Other factors like performance, latency, scalability and availability of legacy systems can all play a role in deciding how to allocate simulation workloads.

What software licensing considerations should users take into consideration when it comes to their data center utilization?

Wim Slagter, Ansys: Ansys offers a variety of licensing types to address typical workloads as well as burst and fluctuating workloads. Besides our traditional perpetual and lease licenses, we also offer usage-based or subscription licenses.

While traditional licenses are most cost effective for long term, consistent usage in an onsite data center, the usage-based licenses offer flexibility that will align better with fluctuating workloads or cloud-based deployments. Our licensing can be used both on-premise or on the cloud; there is no difference.

Rod Mach, TotalCAE: Many users are still license-bound, and to run many CPU or GPU jobs typically there is an additional license cost from the major CAE vendors. A single server today can have 384 CPU cores, and many clients may not have sufficient licenses to use the scale of modern hardware right away. If your company is not investing in cloud-scale licensing or flexible license models the CAE vendors provide to accommodate cloud and hybrid usage, it is often a better investment to get more CAE licensing up to your on-prem capacity, instead of paying a premium to rent cloud machines where you can’t take advantage of cloud scale.

Michael McNerney, Supermicro: Software licensing can be complex depending on the licensing model. An analysis of the lowest TCO while maintaining the service-level agreements to the users is important. Reducing the licensing costs based on simultaneous cores used may save some of the licensing cost, but if the results take too long to be returned to the user then the choices (of the license) will not be acceptable.

Is there anything else you think we should mention in the article?

Michael McNerney, Supermicro: Many organizations will have to evaluate how they want to move forward with their data centers as AI is becoming part of many organizations’ workflows. New generations of servers will require new cooling techniques and increased power delivery to the racks. The redesign needs to be done at the rack level at a minimum, and most likely at the data center level for best results.

Rod Mach, TotalCAE: Today, we see all the major cloud vendors being capacity constrained on the high-end instances used in CAE; there is no guarantee that when you want to access a machine on-demand, there will be capacity given to you. To work around the issue of capacity, many of the cloud vendors for high-end GPU instances now require scheduling that capacity into sometime in the future, which might be acceptable for AI training workloads, but not for CAE simulation requirements that need more deterministic turnaround times to impact designs. Utilizing some space in your existing data center or co-location footprint under your direct control with a hybrid deployment is a low-cost way to ensure you have a minimal capacity to perform the required work when there is no on-demand capacity, or in the event that a major event takes a cloud vendor region offline for an extended period of time.

More Ansys Coverage

More TotalCAE Coverage