In this column, we often explore the trade-offs of performance in this era of multi-core CPUs. A question particularly relevant in the CAD world is how to best balance achievable performance for multi-threaded workloads that benefit from climbing core counts with the ability to still maximize throughput for single-thread execution, still common in CAD workflows. The unfortunate reality underlying the tradeoff is that the top-line CPU specifications that primarily impact performance for the two disparate computing types — CPU frequency and core count — compete with one another, forming an inverse relationship, where one increases as the other declines.
A close examination of that competing interplay between core count, base frequency, and boost frequency is an insightful exercise. Plus, it’s especially important to CAD workflows which are among the most computationally intensive, combining intense demand and a diverse workload.
This month, I’ll recap the impact of the inverse relationship between a CPU’s base frequency and core count, and focus more on the behavior of core count and boost frequency. Boost frequency may be more revealing when it comes to choosing the best combination of CPU metrics for your next CAD machine.
Image source: Quardia, Inc./stock.adobe.com
Push & Pull: Base Frequency and Core Count
First, let’s take a look at the straightforward, yet inverse relationship between core count and base frequency. It’s understandable given the power and thermal (heat) implications of high-performance computing, where component and system designers want to push the performance throttle as close to the edge as possible, without causing system failure. Faster CPU clocks yield higher performance (as long as the rest of the system can keep up). If everything works to an equal power, such as voltage, chip process, and system cooling, faster clocks chew up more power and create more heat. The limits of power delivery and thermal dissipation represent that edge, and exceeding either can crash or even permanently damage the system.
Intel and AMD do all they can to walk that line, optimizing performance without incurring significant failure risk. One way they do so is to judiciously set clock rate limits to achieve best possible performance without short-changing reliability. (It’s worth noting that with chips such as Xeon that are highly reliable workstations and servers, Intel may intentionally dial down spec’d clock rates a bit from their die-equivalent Core-brand siblings to reduce that risk.)
The tricky part is that line typically shifts with core count, as larger, higher core count CPUs face more challenging electrical and thermal constraints than smaller, lower core count CPUs. As a result, the specified base frequencies of CPU SKUs in the same family tend to decline as you increase the core count.
AMD’s Vermeer family, currently bearing the Ryzen 5000 series brand, exemplifies that tradeoff. As the core count rises, the specified base frequency drops. Each SKU’s base frequency indicates the guaranteed minimum frequency the CPU will be able to sustain for all cores running at maximum throughput. Depending on several factors — how many threads are running on how many cores keeping how many core resources busy for how long — your CPU may be able to maintain a higher clip, even for longer periods.
Boost and base CPU clock frequencies for performance-oriented members (Ryzen 7 and Ryzen 9) of AMD’s Ryzen 5000 family (Source: data from AMD).
Pull & Push: Base Clock vs Boost Clock
There’s another inverse relationship in play with respect to CPU frequencies over that same core-count axis that can significantly impact single-thread performance — the correlation between a part’s core count and its boost frequency. You may be familiar with Intel’s Turbo Boost and AMD’s Precision Boost and Turbo Core technologies. While differences exist, the approaches share the same basic concept and goal: to drive the clock frequency for one or more cores beyond the base frequency for as long as allowed by the system environment — namely supplying power and the bigger, associated problem of dissipating the resulting heat. In the context of this column, I will use the generic term boost clock or frequency.
The relationship between base and boost clocks may seem counter-intuitive. At first glance, you might expect that where base clocks rise, boost clocks will follow. Instead, usually where base clocks typically fall at increased core counts, boost clocks often rise. AMD’s Ryzen 5000 is a good example because it spans a sizable range from 8 to 16 cores. While the highest core count SKUs spec the lowest base frequencies, they also spec the highest boost frequencies. For example, the max-core Ryzen 9 5950X integrates 16 cores with the lowest base frequency in the family.
Why the different behavior in base and boost clocks? The reason interestingly stems from the same basic reasons outlined above with respect to core count vs base frequency. As core counts rise, the silicon area within a die and/or multiple dice in a package also rises, and thereby is more likely to limit the chip’s ability to sustain guaranteed base clock rates. But that increased area also means increased mass and surface area — be it represented as more dice, a bigger package, or more interconnection. Now think about that higher core count SKU running a short-duration boost. The bigger mass and surface area (all else being equal) is likely to better tolerate a short-term increase in the thermal output produced by the higher frequency. That doubles if only one or a few cores are currently executing.
Intel’s Individual Core Families
Intel differentiates its boost frequency rates at levels capable for single or all cores. Not surprisingly, the all-core rates tend to follow the base rates — where one is higher, so is the other. That makes sense, because when all cores are loaded, there is no thermal benefit to exploit. For cases where single core execution can raise rates further, Intel’s base and turbo rates demonstrate a similar inverse relationship, as illustrated by the recent 10th Gen Core i7 and i9 SKUs.
When comparing Intel's 10th Gen Core i7 and i9 SKUs, Intel's base and turbo rates demonstrate a similar inverse relationship, as illustrated above.
Why Base vs. Boost Indicates Best CPU for CAD
With all of this in mind, here’s the question for a CAD user: if the SKU with the highest base frequency tends to have a lower boost frequency, and vice-versa, then should we weigh our next CPU purchase on the basis of base or boost frequency?
That brings us to the crux of this two-part column: Exploring how Ryzen 5000 SKUs with varying core counts, base clocks and boost clocks perform on both single-thread and heavily threaded workloads that are prevalent in CAD workflows.
Liquid Cooling Benefits
This analysis actually began out my testing of a Vermeer-class workstation, BOXX’s Apexx A3 Denali, documented a few months back when running our first round of testing. It became obvious during that testing that the BOXX Apexx A3 Denali’s liquid-cooled design is a value-added feature that allows the CPU to hit higher frequencies more often. If it’s more effective at removing heat from the CPU die, then it’s likely better at helping it maintain a boost clock for a longer duration than an air-cooled machine. Given that the vast majority of workstations do not offer liquid cooling, it may not yield results that reflect the behavior most CAD users experience. As such, I set out to test an air-cooled machine that represented a majority of current CAD users. Thanks to Puget Systems and AMD, I got that opportunity.
The System Set Up
Puget Systems, a small-but-mighty supplier of computers out of the Pacific Northwest, built a Ryzen 5000-based X570-E workstation for me, outfitted with a top 16C Ryzen 9 5950X and complemented by top-end memory, storage, and GPU. As with BOXX — and any other workstation vendor not named Dell, HP or Lenovo — a sound business strategy is built on differentiation. With the pricing advantage that volume brings, there’s no point building the same machines, the same way, with the same components and same level of service.
Puget Systems understands this, and the company takes differentiation seriously, most notably in its system build, verification, and customer service. From the moment the machine build order began, the level of attention and service was crystal clear. In fact, the build, delivery and examination of the system so impressed that it’s triggered a topic for a future column. Look for more on Puget Systems and perhaps another small-but-mighty vendor down the road.
Along with the BOXX A3 Denali and Puget Systems X570-E, AMD provided a 12C 5900X to swap into the two systems for comparison purposes. With the two systems and three CPUs (the BOXX came outfitted with an 8C 5800X), I had four systems available, each with a different SKU CPU, backed by the same basic high-performance supporting components. The only notable difference is the Puget Systems Quadro RTX 5000 GPU, but in this case the tests were specifically created to stress the CPU and leave the other components out of the critical path.
Puget Systems’ impressive build of the Ryzen 5000 platform X570-E workstation.
Specifications for the two test systems. (Data from companies.)
The theoretical, engineering-driven relationships between core count, base clock, and boost clock suggest some different ways to think about how we choose CPUs based on those specs and how we expect to stress the CPU with the workloads we rely on most.
Next month, check in for a detailed rundown on test results and analysis of the systems, along with a fresh perspective on how to weigh the specifications for core count, base clock and boost clock when evaluating a CPU for CAD work.
Share This Post