Evaluating The Performance Cost To AMD SEV-SNP On Modern EPYC VMs

Written by Michael Larabel in Cloud on 13 February 2026 at 10:17 AM EST. Page 8 of 8. Add A Comment.

The oneDNN deep learning library saw little impact from SEV-SNP being active for the CVM.

👁 Llama.cpp benchmark with settings of Backend: CPU BLAS, Model: Qwen3-8B-Q8_0, Test: Text Generation 128. 26.04 Devel: Non-CVM was the fastest.

👁 Llama.cpp benchmark with settings of Backend: CPU BLAS, Model: Qwen3-8B-Q8_0, Test: Prompt Processing 512. 26.04 Devel: CVM was the fastest.

👁 Llama.cpp benchmark with settings of Backend: CPU BLAS, Model: Qwen3-8B-Q8_0, Test: Prompt Processing 1024. 26.04 Devel: CVM was the fastest.

For CPU-based AI inferencing with Llama.cpp there was relatively little overhead to leveraging AMD SEV-SNP on this EPYC 9005 Azure instance.

👁 Llama.cpp benchmark with settings of Backend: CPU BLAS, Model: gpt-oss-20b-Q8_0, Test: Prompt Processing 512. 26.04 Devel: Non-CVM was the fastest.

👁 Llama.cpp benchmark with settings of Backend: CPU BLAS, Model: gpt-oss-20b-Q8_0, Test: Prompt Processing 1024. 26.04 Devel: Non-CVM was the fastest.

👁 Llama.cpp benchmark with settings of Backend: CPU BLAS, Model: gpt-oss-20b-Q8_0, Test: Prompt Processing 2048. 26.04 Devel: Non-CVM was the fastest.

👁 Llama.cpp benchmark with settings of Backend: CPU BLAS, Model: Llama-3.1-Tulu-3-8B-Q8_0, Test: Text Generation 128. Ubuntu 24.04: Non-CVM was the fastest.

👁 Llama.cpp benchmark with settings of Backend: CPU BLAS, Model: Llama-3.1-Tulu-3-8B-Q8_0, Test: Prompt Processing 512. 26.04 Devel: CVM 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. 26.04 Devel: CVM was the fastest.

👁 Llama.cpp benchmark with settings of Backend: CPU BLAS, Model: Mistral-7B-Instruct-v0.3-Q8_0, Test: Prompt Processing 1024. 26.04 Devel: CVM was the fastest.

👁 Llama.cpp benchmark with settings of Backend: CPU BLAS, Model: DeepSeek-R1-Distill-Llama-8B-Q8_0, Test: Text Generation 128. Ubuntu 24.04: Non-CVM was the fastest.

👁 Llama.cpp benchmark with settings of Backend: CPU BLAS, Model: granite-3.0-3b-a800m-instruct-Q8_0, Test: Text Generation 128. Ubuntu 24.04: Non-CVM was the fastest.

For those using non-public large language models, inferencing from confidential VMs protected by AMD SEV-SNP allowed for added security with little performance impact. With the Ubuntu 26.04 development build, the SEV-SNP cost was even less than with the current Ubuntu 24.04 LTS release.

👁 Geometric Mean Of All Test Results benchmark with settings of Result Composite, AMD EPYC SEV-SNP VM Benchmark On Ubuntu 24.04, 26.04 Devel. 26.04 Devel: Non-CVM was the fastest.

When taking the geometric mean of nearly 200 benchmarks, the confidential VM backed by AMD SEV-SNP was at 95% the performance of the non-CVM instance. Between Ubuntu 24.04 LTS and Ubuntu 26.04 development the performance cost was similar but Ubuntu 26.04 performed slightly better overall thanks to the newer Linux kernel, GCC 15 compiler, and other software updates.

For database servers and web servers the performance cost of using SEV-SNP was the greatest as expected given the I/O involved. There the performance impact was 10~15% which for many organizations would be an acceptable trade-off for the increased security provided by Secure Encrypted Virtualization with Secure Nested Paging. Outside of those heavy I/O workloads, the AMD SEV-SNP impact on the AMD EPYC 9005 "Turin" series virtual machines was more modest at around 5% or less.

Add A Comment

If you enjoyed this article consider joining Phoronix Premium to view this site ad-free, multi-page articles on a single page, and other benefits. PayPal or Stripe tips are also graciously accepted. Thanks for your support.

Page: 1 2 3 4 5 6 7 8

URL: https://www.phoronix.com/review/amd-epyc-9005-sev-snp/8

⇱ Evaluating The Performance Cost To AMD SEV-SNP On Modern EPYC VMs Review - Phoronix

Evaluating The Performance Cost To AMD SEV-SNP On Modern EPYC VMs