SMT Remains Very Advantageous For 5th Gen AMD EPYC Performance

Written by Michael Larabel in Processors on 6 February 2025 at 10:50 AM EST. Page 5 of 6. 96 Comments.

The Liquid-DSP and srsRAN software defined radio software enjoyed healthy uplift from SMT on AMD EPYC.

👁 NAS Parallel Benchmarks benchmark with settings of Test / Class: BT.C. SMT Disabled was the fastest.

👁 NAS Parallel Benchmarks benchmark with settings of Test / Class: EP.C. SMT Disabled was the fastest.

👁 NAS Parallel Benchmarks benchmark with settings of Test / Class: LU.C. SMT Disabled was the fastest.

👁 NAS Parallel Benchmarks benchmark with settings of Test / Class: SP.B. SMT Disabled was the fastest.

👁 ACES DGEMM benchmark with settings of Sustained Floating-Point Rate. SMT Disabled was the fastest.

Of course, SMT isn't a universal win. With some of the heavier HPC workloads is where disabling SMT can be a performance win. That's why AMD produces an EPYC BIOS tuning guide with various power/performance recommendations. For cases like HPC clusters, AMD recommends disabling SMT in most of those instances. SMT can be easily disabled from the BIOS or even the Linux kernel boot options or offlining sibling threads from within the OS if desired.

👁 NAMD benchmark with settings of Input: ATPase with 327,506 Atoms. SMT Enabled - Default was the fastest.

👁 NAMD benchmark with settings of Input: STMV with 1,066,628 Atoms. SMT Enabled - Default was the fastest.

In some workloads like NAMD, SMT can still pay off.

👁 OpenVINO benchmark with settings of Model: Vehicle Detection FP16-INT8, Device: CPU. SMT Enabled - Default was the fastest.

👁 OpenVINO benchmark with settings of Model: Vehicle Detection FP16-INT8, Device: CPU. SMT Disabled was the fastest.

👁 OpenVINO benchmark with settings of Model: Person Vehicle Bike Detection FP16, Device: CPU. SMT Enabled - Default was the fastest.

👁 OpenVINO benchmark with settings of Model: Person Vehicle Bike Detection FP16, Device: CPU. SMT Disabled was the fastest.

👁 OpenVINO benchmark with settings of Model: Face Detection Retail FP16-INT8, Device: CPU. SMT Enabled - Default was the fastest.

With the OpenVINO AI workloads, having SMT enabled provided a healthy boost to the throughput but at the cost of higher latency.

👁 Timed LLVM Compilation benchmark with settings of Build System: Ninja. SMT Enabled - Default was the fastest.

👁 Timed Node.js Compilation benchmark with settings of Time To Compile. SMT Enabled - Default was the fastest.

👁 Timed Linux Kernel Compilation benchmark with settings of Build: allmodconfig. SMT Enabled - Default was the fastest.