By Bob Cramblitt, communications director for SPEC.
Ever hear this question: What’s the best workstation for product design?
It’s a loaded one, kind of like asking who’s the best jazz trumpeter, what’s the best path to contentment, or when is the best time of day to be creative.
We are conditioned to want an easy answer, thus the workstation benchmarks that seek to provide a few numbers that characterize CAD performance in its entirety, not taking into account variances such as the size and complexity of the dataset, the principal tasks the user is performing, or how the design is affected by downstream operations within the product development cycle.
The folks at the SPEC Graphics and Workstation Performance Group (SPEC/GWPG), entrusted with developing benchmarks for the newest generations of workstations, know it takes much more than a few numbers to properly characterize performance. And, they suspect that most workstation users know as well. It’s why there are more than 25,000 downloads of SPEC/GWPG benchmarks annually.
Simple Enough, Right?
In analyses done by SPEC/GWPG representatives Pradeep Ramineni and Brian Bothwell based on SPEC Application Performance Characterization (SPECapc) benchmark results, CAD modeling is most affected by CPU frequency, as it is normally a single-threaded task or a series of single-threaded tasks.
Interactive graphics performance for model display and manipulation is affected the most by a strong GPU and at least four CPU cores to drive it. Offline rendering and ray tracing benefit from distributing operations among as many available cores as possible, whether they are CPU, GPU or a combination of the two. Multi-threaded CPU and GPU rendering is especially effective for large and highly detailed models, such as aircraft engine assemblies and architectural models with realistic details.
In modern CAD applications, GPU and system memory play a key role in maintaining interactivity, especially when loading, saving or exporting a complex model. Users can expect their applications to grind to a halt when a model is larger than available VRAM and system RAM.
So, simple enough, right? Get the configuration with the CPU frequency and GPU power to fit your needs, and storage to make sure you can load and move files around at an acceptable rate. But not surprisingly, there is a catch to this scenario; several of them actually.
The Many Modeling Paths
Even within well-defined modeling tasks, applications can take different paths, according to Trey Morton, SPECapc chair. These include the type of external referencing of geometry used by the CAD package, support for different types of modeling, and the effect of other product development operations during and after the generation of models.
Many CAD packages rely on externally referenced geometry to fully populate a model and streamline a modeling process. How this functionality is implemented can make a difference in performance. Whether the reference geometry is housed on the desktop system, on a network, or in the cloud also factors into the performance equation.
The type of modeling can make a big difference in what configurations improve or deter performance. CAD packages typically go down different routes to the same destination based on the type of modeling they support, such as parametric modeling, direct modeling, or a hybrid approach.
“The type of modeling being used can make a tremendous difference, along with the particular ways each engine is used,” says Ross Cunniff, chair of the SPEC Graphics Performance Characterization (SPECgpc) subcommittee. “Some parametric surfaces translate very easily to GPU primitives, which leads to higher performance and greater interactivity. Others require significant computation to ‘regen’ into GPU-displayable content; this will require greater CPU power and memory to achieve higher performance levels.”
Modeling, of course, doesn’t exist in a void, especially as the product development cycle moves away from “throw it over the wall” silos to interdependent processes. In these environments, performance is affected both during and after modeling by downstream processes, including PLM/PDM integration, CAE and FEA, photorealistic rendering, management of design iterations, real-time walkthroughs or VR, and storage and retrieval of models.
“Model structure—what parts are grouped together, how each subassembly is modeled via parametric operations, the complexity of the model—is frequently dictated by the needs of FEA, CFD, and downstream CAM operations,” says Cunniff. “In the case of CAM, for example, the post-processing steps to send the subassembly to a CNC milling machine or a 3D printer require certain model configurations to ensure quality and efficiency, and these configurations are not necessarily the same as those needed for interactive modeling.”
The need for increased GPU power is especially essential if users require real-time navigation through complex models for applications such as architectural walk-throughs or VR training for assembly or servicing.
Multi-Threaded Rendering
While heavily multi-threaded CPUs and GPUs generally yield the best performance results for rendering, the type of rendering used is also critical to component choices.
More traditional rasterization processes—taking an image described in a vector graphics format and converting it into a raster image for display on a computer screen—requires less processing power than ray tracing, the method for generating realistic lighting and shadow effects among virtual objects and backgrounds.
High-end effects such as radiosity, which simulates how objects absorb and reflect light in a real-world environment, require global illumination algorithms that further tax computing resources.
“For rasterization and ray tracing, the GPU is king,” says Cunniff. “More processing units and more GPU memory will quickly add up to higher performance. Radiosity brings a CPU component into it because the algorithms require higher levels of cache and much higher clock speeds.”
Be Aware of Trade-Offs
Designers doing modeling and rendering on the same workstation at the same time need to be aware of performance trade-offs, according to Morton.
“GPU-based rendering shifts the burden of rendering from the CPU to the GPU, but you can run into a situation where if a system has one graphics card and you start a GPU render you will have a decrease in functionality for interactive modeling.”
Rather than trying to figure out the myriad trade-offs in performance for modeling versus rendering, it could be best to avoid the “one size fits all approach”, according to Alex Shows, chair of the SPEC Workstation Performance Characterization (SPECwpc) subcommittee.
“A CAD workstation with a high-frequency CPU paired with a high-compute-core GPU would achieve the best blend of modeling and rendering performance, but might be prohibitively expensive,” says Shows. “Product designers might be better served by moving some of their compute and graphics resources to the cloud for easy accessibility and sharing.
“If moving resources to the cloud is untenable, bottlenecks could be avoided with two task-specific workstations: One that maximizes CPU frequency to deliver the interactivity and responsiveness required for modeling and the other configured for rendering efficiency. This kind of division of responsibilities can improve productivity by parallelizing the work of one user over two machines.”
Measure What You Do
Beyond the vicissitudes of how different CAD applications handle operations critical to overall performance, professionals concerned about performance should be aware of their own usage patterns.
Ideally, CAD workstation users would develop their own benchmarks specific to their workloads, the CAD applications they use, and the functionality they rely on the most. But this is difficult and the terrain is constantly shifting, as new versions of applications introduce novel ways of maximizing efficiency.
The next best thing to home-grown performance measurement is to use standardized benchmarks that most closely align with your application and provide the granularity to measure the myriad paths that an application takes. In that vein, there are SPECapc benchmarks based on Creo, NX, and SOLIDWORKS that run on top of those applications. If you want a smaller, self-contained benchmark, the current version of SPECworkstation includes workloads for CATIA, Creo, SOLIDWORKS, NX, and Showcase, as well as open-source datasets for CFD and FEA. If you are interested primarily in graphics performance, the SPECviewperf benchmark (available in Windows and Linux versions) provides workloads (called viewsets) based on CATIA, Creo, SOLIDWORKS and NX.
All of these benchmarks are available to users for free downloading. Vendors of computer-related products and services that are not members of SPEC/GWPG are required to pay a licensing fee.
The Unending Quest for Productivity
In today’s product development environment, increased integration among CAD, CAE and CAM requires finely tuned choreography. Pauses or hiccups waste precious engineering time.
Users don’t need the biggest or fastest of every component, but will benefit from upgrades that address their most problematic bottlenecks. Benchmarks that deliver alignment with an engineer’s day-to-day work can help organizations make the decisions that will guarantee a better ROI.
Whatever way you decide to benchmark, be aware that performance is a perpetually moving target, affected by frequent updates from CAD application vendors, component manufacturers, driver writers, and OEMs. There are also the continuous demands on product designers to do more in less time, leading to larger datasets that put a greater strain on driver, processing, rendering and storage components.
It’s a commitment to keep up, but the benefit comes from making sure that one of an engineering company’s most valuable assets—its product designers—are equipped with the best tools to be as productive as possible.
_____
Bob Cramblitt is communications director for SPEC. He writes frequently about performance issues and digital design, engineering and manufacturing technologies.