And naturally, graphics cards and the GPUs that power them are not limited to desktop or mobile workstations. They’re also taking over the data center. Engineering organizations are making use of data center GPUs to empower their employees with best-in-class hardware for photorealistic ray tracing, machine learning training and inferencing, engineering simulation and more.
Graphics Cards in the Data Center
The optimal hardware deployment will be different for every company, but there are some very compelling reasons to put graphics cards in a data center instead of a desktop. A high-end desktop workstation for engineers could have one or two high-end graphics cards such as the NVIDIA RTX 6000 Ada Generation—maybe three or four if it was really decked out. But there’s only so far a desktop workstation can go. The system is ultimately limited by its power supply and expansion slots, as well as its acoustics (there’s only so much fan noise an engineer can take, after all).
The data center does away with these limitations. They’re built with maximum power and cooling requirements, unlocking a much higher density of compute power than that of a desktop workstation.
“It is not unusual to see a single datacenter GPU server house 8 to 10 GPUs in a 4U configuration,” says Jay Chen, senior manager of data center and AI solutions at PNY.
Three major benefits accompany this approach. First, data center hardware is easier to manage from an IT perspective than a fleet of high-end desktop workstations. Second, it provides much more flexibility. Rather than each engineer in a company having their own dedicated graphics hardware, data center GPUs allow engineers to share resources as needed.
And perhaps the biggest benefit of data center GPUs is their scalability. “It scales almost linearly as you add more of these GPUs in the server,” says Himanshu Iyer, senior product marketing manager at NVIDIA.
That scalability is key for engineering companies, especially those that need to render extremely large models. Iyer shared that French automaker Renault uses data center GPUs to dramatically decrease photorealistic rendering times and shorten design loops, as described in a March 2022 presentation from NVIDIA’s GTC conference called Next Challenges in Computer Graphics for Automotive Lighting Simulation.
Iyer also pointed to Heesen Yachts, a Dutch shipmaker that takes advantage of data center GPU scaling to dramatically cut down on rendering times while designing its luxury superyachts. “They’re using [Dassault Systèmes CATIA] live rendering, and that is another very good application where we have seen excellent scaling on these data center GPUs,” Iyer says.
Different Types of Data Center GPUs
As the applications of GPUs have broadened, chipmakers have begun designing different types of GPUs for different tasks. NVIDIA, for example, currently offers graphics cards built on one of two concurrent GPU microarchitectures, Hopper or Ada Lovelace. Though these architectures share many similarities, they are designed for different use cases.
“NVIDIA’s Hopper architecture is designed for ultra-high performance data computation,” Chen explains. An example of a Hopper GPU is the NVIDIA H100 for data centers, which Chen says excels at AI tasks such as training large language models. Part of its power comes from a high-speed PCIe 5.0 interface and HBM2e memory that enables up to two terabytes per second of memory bandwidth, essential for handling large AI models.
In contrast, Chen describes the Ada Lovelace architecture as a “universal GPU accelerator.” He says that while Ada GPUs are well-suited to AI training and inferencing—like Hopper GPUs—they also include specialized RT Cores to support AI-assisted real-time ray tracing. In that sense, he says, cards like the NVIDIA RTX 6000 Ada Generation desktop graphics card or NVIDIA L40 data center graphics card can “deliver the full range of capabilities supported by NVIDIA RTX technology.”
“ISV software suites take advantage of NVIDIA RTX technologies like ray tracing, AI and physically accurate simulation to deliver significantly better performance and features when deployed on Ada Lovelace architecture-based GPUs including the NVIDIA RTX 6000 Ada or L40,” Chen says.
According to NVIDIA’s Anthony Larijani, product marketing lead for data center GPUs, one of the most compelling applications for both cards is NVIDIA’s Omniverse software platform for real-time 3D simulation and collaboration.
“It is really a killer use case for all of the enhancements that were made within the Ada architecture, and then running on our highest performance RTX-enabled GPU, the L40,” Larijani says. “So you’re able to work with very, very large scenes and simulations and models that are running in Omniverse.”
The RTX 6000 Ada Generation and L40 are nearly identical cards, spec-wise—other than that, the former is for desktop workstations and the latter is for data centers (one of the biggest differences is that the desktop card has active cooling while the data center card does not). Engineers impressed by the RTX 6000 Ada can consider the L40 as its data center equivalent, and vice versa.
It goes to show that desktop and data center GPUs are simply two different approaches to realizing the many benefits that graphics cards can offer engineers.
“The NVIDIA L40 or RTX 6000 Ada Generation are both very capable of performing diverse GPU accelerated tasks, ranging from CAD and CAE (or other) simulation, real-time ray-traced rendering, or AI and ML development and deployment,” Chen says.