More CPU Cores or Faster CPU Clocks?

1 Dec, 2015 By: Alex Herrera

To best address the demands of a modern CAD workflow, look for a balance of CPU core count and clock rate in your next CAD workstation.

A Case for Maximum Clock Rate

Despite today's overarching focus on multicore architectures, processor vendors such as Intel haven't given up on improving single-thread performance. But rather than looking for, say, a 50% improvement in single-thread execution moving from one generation to the next, designers are targeting an estimated 10% improvement per core, then grabbing another performance bump with the addition of more cores.

CPU vendors continue to pay attention to single-thread performance for a good reason, one that is especially evident in typical CAD computing. Though several common CAD algorithms are able to effectively execute multiple threads on multiple cores, as the Rodinia test case illustrates, there remain critical tasks in CAD workflows that simply aren't. On the contrary, they remain primarily single-threaded in nature, making the case for a single core that can execute at the fastest possible clock rate.

The solving process in the parametric modeling that forms the foundation of most CAD workflows is fundamentally linear in nature, and therefore offers few opportunities to cut execution times with multiple threads.

Another example is the GPU-driven 3D rendering we rely on most often in interactive design work (not to be confused with software-based rendering, which does run well on multiple CPUs). Running Cadalyst's AutoCAD-based benchmark test (C2015), we see the vast majority of work being performed by one core, resulting in a low utilization — about 3% — when considering all available cycles from all cores (figure 3). Other GPU-focused rendering tasks, such as the SPECviewperf viewset for CATIA, show a similar CPU usage profile.

Figure 3. Rendering via GPU-driven 3D graphics has historically not benefited from multicore CPU use.
Figure 3. Rendering via GPU-driven 3D graphics has historically not benefited from multicore CPU use.

For such all-too-common CAD tasks, single-thread performance matters, and the simplest way to increase that performance is to turn up the clock rate. Intel Xeon processors, targeting high-reliability workstations and servers, tend to run up to about 3.7 GHz, while the highest-clocked Intel Core i7 brand processors push a bit higher, maxing out around 4.5 GHz.

Consider also that Intel's specified clock rates for these products are nominal, and Intel does allow vendors to drive Core-brand CPUs to even higher rates. While most high-volume vendors such as HP, Dell, and Lenovo tend to shy away from overclocking, smaller suppliers such as BOXX, Maingear, @Xi Computer, and Digital Storm do overclock CPUs to further accelerate single thread–focused tasks by another 10% or so.

CAD Caveats

Now that we understand different performance profiles of a modern processor in the context of CAD workloads with varying degrees of threading, let's touch on a few caveats. First, every user's situation comprises a unique set of workloads and emphases on tasks — and while it's reasonable to present a typical case for CAD in general, how well a particular combination of core count and frequency will handle each user will vary.

Second, achievable CPU utilization is improving steadily as computing platforms continue to progress, with subsequent generations of hardware, applications, and supporting software. For example, remember the aforementioned reference to how GPU rendering does not currently leverage multiple cores effectively? Well, that will start changing now. The DirectX 12 application programming interface (API) released in Windows 10, for example, is the first version of the API to effectively leverage multiple CPU cores to help ensure GPU-driven rendering won't be throttled by the CPU (figure 4).

Figure 4. DirectX 12, shipping in Windows 10, is the first DirectX API to effectively leverage multiple CPU cores for GPU-rendered graphics. Image courtesy of AMD.
Figure 4. DirectX 12, shipping in Windows 10, is the first DirectX API to effectively leverage multiple CPU cores for GPU-rendered graphics. Image courtesy of AMD.

Third, and perhaps most importantly, the CPU isn't the only performance-critical component in the system, and throughput can often be throttled somewhere else — graphics, memory, or storage, for example. Common tasks such as performing a walkthrough of a virtual skyscraper or rotating and zooming around a 3D model of an engine are just as likely to choke your GPU as your CPU, if not more so. Similarly, reading model data from disk is something that multiple cores could team up on, but the read rate of the drive (or drive interface) would most likely limit bandwidth far below what the CPU cores could handle.

In those cases where the CPU isn't the bottleneck, it probably won't matter whether one CPU with more cores is running the application more efficiently than another CPU with higher frequency, or vice versa. For example, check out the CPU utilization running another numerical solver, Calculix (figure 5), used primarily for finite-element analysis (FEA). We still see very effective multithreading, with all cores similarly busy (at least averaged over time). But utilization is nowhere near the 100% of the previously introduced, best-case Rodinia simulation. It's instead closer to 50%, as performance is being throttled elsewhere in the system; in this case, it appears to be memory.

Figure 5. The FEA numerical solver Calculix test case shows effective multithreading, but with performance throttled elsewhere in system.
Figure 5. The FEA numerical solver Calculix test case shows effective multithreading, but with performance throttled elsewhere in system.

Another Good Argument for More Cores: You

Having said all that, it's worth considering that the most effective multitasking component in today's CAD workflow is you, the user. Outfit a CAD professional with several high-resolution displays to help manage the workflow and invariably, he or she will find ways to juggle multiple computing jobs — perhaps modeling in one window, while running a simulation in another, and using a third for office applications, browsers, and e-mail. The fact that several are running in parallel takes advantage of multiple computing cores in the system to improve the ultimate metric that matters: productivity.

The Best Answer Is Balance

At this point, I reach two basic conclusions. First, a modern CAD workflow will include some tasks that run better on one or two high-frequency cores, along with others that perform better on a larger array of cores with more modest frequency. And second, because of this, it's simply not possible to choose a CPU that will provide the best performance under all possible circumstances.

Still, while it's hardly a perfect exercise, it's worth spending time to think about which CPU is best equipped to run the various types of code and workloads you most frequently run, and, of course, what your budget allows.

Here's where we get to that bottom line of balance: With common CAD tasks representing both efficiently multithreaded code and fundamentally single-threaded code, and with a host of other variables in the equation dictating overall, realized system-level performance, the only reasonable choice is to select a CPU that achieves a good balance of core frequency and core count without breaking the bank. To make the selection process more manageable, you'll find Intel and workstation developers have already filtered reasonable Intel Core and Intel Xeon options down to a handful, itemized on any vendor's datasheet or online configuration tool.

Today, four cores (quad-core) is a sensible lower limit, and eight a reasonable upper limit. If you find your time is almost exclusively spent modeling with small to mid-size models, then a quad-core CPU with a higher frequency would be a fine choice.

You might even consider an overclocked, liquid-cooled model. And, if you find yourself often waiting for simulations, particularly with larger, more complex models, then a higher–core count CPU could be a wiser option. Either way, paying attention to the CPU options a workstation vendor offers, considered in the context of your own workflow, will pay long-term dividends by helping you finish the job in the shortest time possible.

1 2 

About the Author: Alex Herrera

Alex Herrera

Add comment

Note: Comments are moderated and will appear live after approval by the site moderator.

AutoCAD Tips!

Lynn Allen

In her easy-to-follow, friendly style, long-time Cadalyst contributing editor and Autodesk Technical Evangelist Lynn Allen guides you through a new feature or time-saving trick in every episode of her popular AutoCAD video tips. Subscribe to the free Cadalyst Video Picks newsletter and we'll notify you every time a new video tip is published. All exclusively from Cadalyst!

Follow Lynn on TwitterFollow Lynn on Twitter

Download Cadalyst Magazine Special Edition