A valuable feature that can help boost performance and efficiency (and at no additional cost**), offered across nearly all AMD EPYC™ processors, is simultaneous multithreading (SMT), which allows a single physical CPU core to execute multiple threads simultaneously. As shown in Figure 1, this is achieved by allowing the operating system and applications to utilize two logical processors for each physical core, and sharing many of the core's computational resources, enabling it to work on two tasks at once. If one thread stalls due to a cache miss while waiting for data, the other can seamlessly proceed, ensuring that precious processor core resources are minimally idle.
Figure 1: Overview of SMT Concept
The SMT feature in AMD EPYC processors may be toggled on or off as needed, which is valuable as some workloads, such as high-performance computing (HPC), may perform better without it, while other workloads realize much better performance uplift by utilizing multiple threads.
What is SMT and Why Does It Matter?
Since its creation, SMT has become a staple of enterprise-class processors as a proven method to get as much out of the investments in silicon as possible. AMD included SMT with the launch of Ryzen and EPYC CPUs, and the feature has been incorporated into virtually all AMD EPYC processors since. This has provided a powerful complement to the leadership number of physical x86 cores that AMD has provided with each generation of EPYC—with the recent 5th generation EPYC 9005 family topping out at a whopping 192 physical cores and 384 threads per socket! In a dual socket system, that is a LOT of computing power in a 1U or 2U system!
Intel had also offered SMT capabilities in its Xeon server processor line for many years, and this was critical in recent times when Intel was constrained in its ability to provide more physical cores in each CPU generation. In an interesting contrast, the recent generations of Arm-architecture-based processors, including those from Ampere Computing, Nvidia, and hyperscale cloud providers, lack SMT capabilities. This is most interesting, since an oft-cited “advantage” of these processors is high levels of efficiency—although it is well established that there are MANY workloads that perform better and more efficiently with SMT enabled—even in systems equipped with a very large number of physical cores! One of the more comprehensive such analyses was published in 4Q’24 by the testing experts at Phoronix, including these closing comments:
“For workloads able to benefit from SMT, it's still a clear win with AMD EPYC 9005 processors. When looking at all of the CPU power consumption across 170+ benchmarks taking ~13 hours to complete, the data here shows no power consumption difference overall to having SMT enabled.”
Source: Exploring The Zen 5 SMT Performance With The AMD EPYC 9755 "Turin" CPU - Phoronix
The broad survey of workloads shows that the performance impact will vary by workload. In the wider set of workloads, SMT provides real value, so administrators can turn that function on easily. Some high-performance computing (HPC) applications, which often utilize threads that demand exclusive access to critical core resources, may see little-to-no advantage from SMT. In these cases, administrators can easily disable SMT, tailoring their systems for optimal performance. This ability to fine-tune hardware behavior makes AMD EPYC processors a compelling choice for enterprises with diverse IT needs.
It seems clear that a vendor interested in allowing customers to maximize their efficiency would want to enable SMT, but they chose not to. But in another curious development, Intel appears to have abandoned its use of SMT with the release of “Sierra Forest” processors with “efficient cores” (e-core), part of its Xeon 6 lineup, in 2024. While this would seem to allow Intel to compete more directly with the Arm options, it does so on terms that seem to limit customer choice while also creating greater divergence among Intel’s Xeon portfolio—as prior generation Xeon and the latest “Granite Rapids” Xeon 6 parts with “P-cores” will support SMT.
For IT buyers, AMD EPYC and its consistent deployment of the SMT feature can provide an advantage when selecting platforms for their workloads. By dynamically managing threads and allocating resources efficiently, AMD has ensured that workloads ranging from enterprise databases to random-access-heavy applications, such as AI training, can benefit significantly from SMT. AMD processors give users the flexibility to enable or disable the feature based on specific workload requirements, a versatility that sets it apart.
SMT: Performance Boost at No Additional Cost
In many scenarios, SMT acts like a free turbocharger for processors, doubling the thread count per core. For workloads optimized to leverage multiple threads—such as databases, virtualization, and certain enterprise applications—this can result in performance gains of 30-60% without the need for additional physical cores. Sure, this allows for some savings in hardware—perhaps buying less costly CPUs with fewer cores, but there is so much more. This is particularly beneficial for organizations with enterprise software that is licensed based on physical core count metrics (examples include, Microsoft, Oracle), as they can achieve better throughput on fewer cores and therefore minimize licensing costs.
As one can see, there is a notable performance upside enabled by SMT. And when the software license price is thousands or tens of thousands of dollars per CPU core used—for each of dozens or hundreds of servers - the savings opportunity could be quite significant.
Cost and Energy Efficiency
By enabling more threads to run on fewer physical cores, SMT helps organizations consolidate workloads, reducing the number of servers required. This can further translate into lower capital expenditures and reduced energy consumption—a critical consideration as sustainability becomes a priority for data centers.
While Arm advocates argue that removing SMT reduces power consumption and chip complexity, the savings are minimal compared to the performance gains SMT delivers and potential savings in licensing and other costs. And Intel’s inconsistent approach – offering hyperthreading in some CPUs, but not others – reflects architectural trade-offs that limit flexibility and seamless portability for users.
AMD has maintained its focus on refining SMT with each processor generation, ensuring that power and silicon area overheads are minimized, while maintaining high performance per watt and performance per CPU dollar value. This balance of efficiency and capability makes SMT a highly valuable feature for enterprise IT environments.
Tailoring Infrastructure for Maximum Value
For enterprises weighing the benefits of on-premises infrastructure versus cloud deployments, SMT provides a compelling argument for maintaining control. Cloud service providers often standardize instances, offering T-shirt-sized options that bundle physical and virtual cores without user customization. This lack of flexibility can lead to inefficiencies, as enterprises may end up paying for unused resources. Since the story for the cloud can be a bit muddier, we’ll offer a blog to look at the cloud situation regarding SMT and other configuration attributes in short order.
On-premises deployments with AMD EPYC processors allow IT teams to tailor configurations to specific workloads, optimizing cost and performance. By enabling or disabling SMT as needed, organizations can ensure they’re getting the best value for their investment.
SMT is a Strategic Advantage – Demand It!
In the race to deliver high-performance, efficient IT infrastructure, SMT represents an advantage for AMD. Its ability to boost performance without additional hardware, combined with flexibility and cost-efficiency, makes it a powerful tool for IT buyers navigating complex workload demands.
For enterprises seeking to future-proof and optimize their infrastructure, AMD’s EPYC processors with SMT offer a clear path to balancing performance, cost, and flexibility—ensuring they’re equipped to handle the challenges of today’s data-driven world.
** Note that several leading cloud services providers market cloud services based on “instances”, “machine types” or other units of measure that are sized and priced based on a mix of physical and virtual cores enabled by SMT.