Measuring The Performance Cost To AMD Memory Guard With Ryzen AI PRO CPUs

Written by Michael Larabel in Software on 12 June 2025 at 08:00 AM EDT. Page 2 of 4. 11 Comments.

👁 OpenRadioss benchmark with settings of Model: Rubber O-Ring Seal Installation. DRAM Encryption Disabled was the fastest.

👁 OpenRadioss benchmark with settings of Model: INIVOL and Fluid Structure Interaction Drop Container. DRAM Encryption Disabled was the fastest.

The software with the biggest performance hit from having full DRAM encryption was Altair's OpenRadioss, the memory-intensive FEA software. Disabling memory encryption improved performance by just under 10%.

👁 DaCapo Benchmark benchmark with settings of Java Test: H2 Database Engine. DRAM Encryption Disabled was the fastest.

👁 Renaissance benchmark with settings of Test: In-Memory Database Shootout. DRAM Encryption Disabled was the fastest.

👁 DaCapo Benchmark benchmark with settings of Java Test: Tradebeans. DRAM Encryption Disabled was the fastest.

Some memory heavy OpenJDK Java workloads saw several percentage performance hits from this default DRAM memory encryption via Memory Guard.

👁 LocalScore benchmark with settings of Model: Llama-3.2-1B-Instruct-Q4_K_M.gguf, Acceleration: CPU. DRAM Encryption Disabled was the fastest.

👁 LocalScore benchmark with settings of Model: Meta-Llama-3.1-8B-Instruct-Q4_K_M.gguf, Acceleration: CPU. DRAM Encryption Disabled was the fastest.

👁 LocalScore benchmark with settings of Model: Qwen2.5-14B-Instruct-Q4_K_M.gguf, Acceleration: CPU. DRAM Encryption Disabled was the fastest.

👁 LocalScore benchmark with settings of Model: Llama-3.2-1B-Instruct-Q4_K_M.gguf, Acceleration: CPU. DRAM Encryption Disabled was the fastest.

The LocalScore Llama.cpp-powered AI benchmark also clocked in some measurable performance hits as a result of the Memory Guard RAM encryption.

👁 Llama.cpp benchmark with settings of Backend: CPU BLAS, Model: granite-3.0-3b-a800m-instruct-Q8_0, Test: Prompt Processing 1024. DRAM Encryption Disabled was the fastest.

👁 Llama.cpp benchmark with settings of Backend: CPU BLAS, Model: Llama-3.1-Tulu-3-8B-Q8_0, Test: Prompt Processing 1024. Full DRAM Encryption - Default was the fastest.

👁 OpenVINO benchmark with settings of Model: Noise Suppression Poconet-Like FP16, Device: CPU. Full DRAM Encryption - Default was the fastest.

👁 OpenVINO benchmark with settings of Model: Face Detection FP16, Device: CPU. DRAM Encryption Disabled was the fastest.

Similarly, other AI workloads like OpenVINO and Llama.cpp clocked some performance costs to AMD Memory Guard given the memory intensity of AI inferencing.

👁 nginx benchmark with settings of Connections: 500. DRAM Encryption Disabled was the fastest.

👁 nginx benchmark with settings of Connections: 1000. DRAM Encryption Disabled was the fastest.

Memcached

Memcached another workload seeing a measurable hit to performance. Thankfully these workloads aren't too common for laptops, but they can be for edge / SOHO server type systems and in those edge deployments the security can be quite important.

OpenAPV

Some video encode workloads were also affected due to the AMD Memory Guard feature by default with the Ryzen AI Max+ PRO 395.

11 Comments - Next Page

Page: 1 2 3 4 Next Page

URL: https://www.phoronix.com/review/amd-memory-guard-ram-encrypt/2

⇱ Measuring The Performance Cost To AMD Memory Guard With Ryzen AI PRO CPUs Review - Phoronix

Measuring The Performance Cost To AMD Memory Guard With Ryzen AI PRO CPUs