Simultaneous Multithreading: Boosting Performance and Efficiency on AMD EPYC Processors

Uncompromised Performance: Maximizing Efficiency with AMD EPYC and SMT

One of the most beneficial things for the Arm-based processors relates to the predictability of performance and energy efficiency. To achieve such a level of efficiency and performance, vendors created the possibility to function on several threads concurrently. This technology is used in enterprise-class CPUs and is called SMT.

Arm vendors share ideas that because of SMT, arises unpredictability of performance, occur security vulnerabilities, and additional costs are needed for the implementation of this technology. Despite such a reaction, ARM supports multi-threading in embedded uses and specific processors. Because of the various contradicting ideas from the vendors, let’s discuss this topic more to get a better understanding of simultaneous multithreading

Understanding SMT

SMT is a technology that was designed to allow the CPU core to deal with several threads at the same time. Once this technology appeared, lots of processors used it for various numbers of threads. The most standard approach that we will discuss in the article is 2-way – it functions by executing 2 threats at the same time.

Advantages of Leveraging SMT

• Better throughput. The usage of 2 threads simultaneously definitely influences improved performance and increased IPC or Instructions Per Cycle.
• Better resource usage within the core. Because of SMT usage, it is possible to create such a process when the core is dynamically busy all the time. When there is a gap in the functioning of the first thread because of the waiting of the necessary data, the second continues to function and use core resources.
• Efficiency of the energy usage. Because of SMT, it is possible to raise the performance, but at the same time, it doesn’t influence the power consumption.
• A possibility to significantly improve performance without the necessity to add license costs.
• Flexibility. SMT might be disabled or enabled via BIOS, and this creates a great flexibility for the admins to initiate the ideal workload needed.
• Software support. SMT has been used for more than 20 years already, and during this period, software support has reached that high level where most modern ecosystems fully support SMT. This feature works out of the box in most case scenarios, so developers can use it right away without the necessity to additionally support it.

Design Challenges in SMT Implementation

• The increased surface of the attack. Attacks are so frequent nowadays and can access absolutely anything; that’s why any feature is considered as a potential surface attack. There are lots of tests relating to possible vulnerabilities specifically because SMT is a potential target for attacks. Over the past 20 years, vendors have greatly minimized the possibilities of such threats. AMD used such security feature as SEV-SNP, and the process of detecting new threats still continues.  
• One more possible issue is the challenge in proving excellent performance for both threads. It is crucial to determine how to effectively schedule instructions and which resources should be shared. To support the principle of equal sharing, there should be some control mechanisms implemented.

The Cost of SMT: Is It Worth the Investment?

From the perspective of the common user, it is just a standard feature that is already there and doesn’t need any additional investments. It can be easily turned off and on whenever it is necessary.

When talking about the topic of spending in terms of feature implementing, the price is relatively small, and its benefits outweigh minimal investments that are done. For instance, the implementation of SMT takes even less than 5 percent of the core area when taking into consideration the latest cores. This relates to all the important logic to help with the core resource sharing. That means SMT uses minimal silicon area and provides exceptional results of functioning.

Moreover, in case the software is licensed depending on the number of cores, extra performance can directly impact the cost saving.  

How SMT Enhances Performance and Efficiency

AMD EPYC processors have already proven their efficiency and high-performance characteristics. That also relates to the workloads that benefit from SMT and multithreading, and there are also those that don’t, such as computing and HPC apps.

The test that was conducted by Phoronix on Zen 5 showed extremely improved performance in most workloads testing, specifically cryptography and databases.

Also, it is important to mention that tests on the 5th and 4th generation of EPYC with the different workloads showed almost no or minimal difference in the power consumption when the feature is disabled vs enabled.

Average testing showed that SMT on EPYC 9575F had the increasement of only 2 watts when compared with the disabled feature.

The SMT is really efficient, and we will once again try to prove that fact. When the core is executing instruction, thread may stop and wait for data, and during waiting process the usage of power is the same, when the second thread is added, the gad is immediately filled and the performance is improved. That means power efficiency is way more increased than a slight increase in the power consumption.  

AMD EPYC with SMT: Consistently Delivering Exceptional Value

SMT technology was developed a long time ago when 2 or 4 cores per socket were used, and everybody was trying to squeeze the maximum out of the available resource. Right now, the situation is way better because EPYC processors have up to 192 cores per socket, but the SMT technology still continues to be effective. This feature is definitely impactful and can balance the demand for more resources and necessary investments. SMT is a fantastic choice to boost the performance and do it almost for free of change.