SMT Performance Comparison: Xeon 678X vs Threadripper 9980X

In reporting by Phoronix’s Michael Larabel, a test comparing the effects of simultaneous multithreading on current workstation processors from Intel and AMD has been published. It was published on September 17, 2026. The subjects are Intel’s Xeon 678X and AMD’s Ryzen Threadripper 9980X. The former belongs to the Granite Rapids WS generation, the latter to the Zen 5 generation. It is notable for comparing Hyper-Threading and SMT, the two vendors’ threading technologies, in both enabled and disabled states.

As a follow-up to the article earlier this month of the Intel Xeon 600 workstation performance against AMD Ryzen Threadripper 9000 series in over 400 benchmarks – including when carried out at the same core count and memory configuration

As the above quote indicates, this test follows the comparison of Xeon 600 versus the Threadripper 9000 series in more than 400 benchmarks published earlier this month. The previous test included conditions with matched core counts and memory configurations. This time it is positioned as a sequel focused on the presence or absence of simultaneous multithreading. It was conducted as the final phase of loaner testing of the Xeon 678X on an HP Z4 G6i.

Workstation Configuration and Test Conditions

The Intel test system used for validation was an HP Z4 G6i workstation. The installed Xeon 678X has a standard configuration of 48 cores and 96 threads. In addition to that configuration, it was also measured with Hyper-Threading disabled at 48 cores and 48 threads. The aim is to separate physical-core performance from the added gain of extra logical threads.

The AMD test system was a System76 Thelio Major workstation. The installed Ryzen Threadripper 9980X has a standard configuration of 64 cores and 128 threads. In addition to that configuration, it was also measured with SMT disabled at 64 cores and 64 threads. Furthermore, a configuration was prepared using BIOS settings to enable only 6 cores per compute die, for a total of 48 cores. This is a measure to enable a comparison with matched physical core counts against the 48 cores of the Xeon 678X.

The core microarchitecture is Redwood Cove on the Intel side and Zen 5 on the AMD side. The two differ in design philosophy and in how front-end resources are shared. Even for the same workload, the gains from simultaneous multithreading may differ. The focus of this test is to verify that difference with real workstation workloads.

Purpose of the Equal Core-Count Comparison

Comparing with matched core counts serves to eliminate simple competition in installed core numbers. If a 64-core part and a 48-core part are compared as-is, the difference in parallelism will dominate the results. Matching at 48 cores highlights per-core processing capability and the efficiency of simultaneous multithreading. This is seen as highly reproducible as material for procurement decisions.

The previous test with more than 400 benchmarks was important in that it also matched memory configurations. If differences in bandwidth or latency contaminate the results, the evaluation of the processors themselves is distorted. This time, as an extension of that, it isolates only the effect of thread count. The staged testing procedure can be judged reasonable.

Intel’s runtime environment improvements are continuous with the developments covered in Intel PISA Adoption, LEO Default on Nova Lake. A processor’s capability is not determined by standalone performance alone. Effective performance changes with the combination of execution infrastructure and operational settings. A perspective that treats SMT enablement as part of operational settings is considered essential.

Operating Environment and Measurement Scope

in Detail

The base OS was Fedora Workstation 44. The kernel was Linux 7.1 and the compiler toolchain was GCC 16.2. All belong to new generations at the time of testing. It can be evaluated as a configuration that readily reflects new instructions and power-management improvements.

The measurement targets are workloads where the trade-offs of simultaneous multithreading are likely to appear. These are seen to include rendering, video encoding, compilation, scientific computing, and compression. In workloads prone to resource contention, adding logical threads can be counterproductive. Conversely, in workloads with long wait times, overall throughput can increase.

Power consumption and heat output were also observed for reference. Processor temperature and power consumption are directly linked to sustained performance. Short-term peak speeds and long-term stable speeds often do not match. For workstation systems, the latter stability is seen as important.

Impact of SMT On/Off on Real-World Workloads

The impact of SMT on real-world workloads is not uniform. In easily parallelized rendering and video processing, adding logical threads can boost throughput. In workloads with intense branch-prediction or shared-cache contention, gains can slow. Understanding the characteristics of each workload is seen as a prerequisite for operation.

As an example of workloads that benefit from parallelization, the design discussion of agent execution frameworks organized in LangGraph vs AutoGen vs CrewAI: AI Agent Comparison 2026 is a useful reference. In configurations that run multiple agent processes concurrently, how thread resources are used affects overall responsiveness. Enabling or disabling simultaneous multithreading acts as a control valve for response latency and throughput. Developers are seen as needing to use them selectively according to the wait characteristics of workloads.

In compilation workloads and continuous-integration queues, enabling simultaneous multithreading is seen to often improve overall efficiency. On the other hand, in latency-sensitive interactive processing, securing dedicated cores can contribute to stability. It should be noted that uniform enablement is not the optimal solution. Establishing guidelines by use case can be evaluated as determining on-site productivity.

Significance of Power Consumption and Thermal

Observations

Observing power consumption and heat output is highly significant. Adding logical threads increases processing volume per unit time while also increasing power consumption. From a performance-per-watt perspective, there are workloads where enablement is not advantageous. If frequencies are limited by thermal caps, single-thread performance can conversely decline.

Control of power consumption and heat overlaps with the importance of thermal management shown in Intel Thermald 2.5.12 Adds Initial ARM Support. Kernel and user-space thermal control mechanisms dynamically adjust frequency and power distribution. The presence or absence of simultaneous multithreading changes the preconditions for that adjustment. Temperature trends during measurement are seen as also affecting quietness and chassis design.

Workstation systems assume prolonged high loads. Sustainable frequencies, rather than momentary peak values, determine completion times for deliverables. Power and thermal data are also directly linked to total cost of ownership considerations at procurement time. Reporting performance and power measurements together can be evaluated as having high practical value.

Editorial Opinion

This SMT test is seen as prompting a review of procurement criteria within 3 to 6 months. Comparisons based only on thread-count figures will recede, and enable/disable guidelines by workload will be established. Vendors are also seen as clarifying recommended settings by use case. In 1 to 3 years, balancing increasing core counts with power constraints is seen as becoming central to design. For workloads where simultaneous multithreading gains are thin, power-saving operation will be chosen. A power-saving perspective is evaluated as becoming established in operational management and development guidelines. The remaining questions may be sharing criteria for which workloads should have simultaneous multithreading disabled. Is the reproducibility of the published test conditions sufficient? Is the disclosure of power and thermal data at a level sufficient for procurement decisions?

References

• “Intel Granite Rapids WS vs. AMD Threadripper Zen 5 SMT Performance”, by Michael Larabel — Phoronix, 2026-09-17T19:14:14.000Z (ARR)
• Source URL: https://www.phoronix.com/review/amd-zen5-intel-gnr-ws-smt