Performance through the customer lens
Security is the reason organizations turn to confidential computing. They need assurance that data remains protected while it is being processed, supported by attestation and a clearly defined root of trust. Once that foundation is understood, attention quickly shifts to a second consideration: how the protected environment performs.
For customers, performance is not an abstract benchmark. It affects application throughput, response time, capacity planning, and ultimately the efficiency of the infrastructure they operate. The practical concern is whether stronger data-in-use protection changes those outcomes, and whether newer platforms are narrowing any difference.
Performance data for Azure Confidential VMs powered by AMD EPYC TM processors helps answer that concern from two complementary angles. The useful story is not a single percentage labeled as confidential-computing overhead. It is how the platform has advanced across generations and how the current confidential offering performs against an appropriately matched general-purpose baseline.
- What does a customer gain by moving from the previous AMD confidential VM generation to the current one?
- Within the current processor generation, how much of the matched general-purpose VM performance does the confidential VM preserve?
Each comparison answers a different customer question. Treating them separately prevents processor-generation improvements from being mistaken for the cost of confidentiality.
Reading the data through two lenses
Lens 1: progress across confidential VM generations
The first lens tracks how the confidential platform itself has advanced. For AMD-powered Azure Confidential VMs, which use the SEV-SNP variant of AMD Secure Encrypted Virtualization (SEV, AMD's umbrella term for its VM memory encryption and isolation technology), it compares DCas_v6, based on 4th Generation AMD EPYCTM processors, with DCas_v5 from the prior processor generation, using independent results from Prowess Consulting. It also previews the next step, comparing DCas_v7 on 5th Generation AMD EPYCTM processors with DCas_v6, based on Microsoft internal testing ahead of the DCas_v7 preview. Because both sides of each comparison use confidential VMs, the result reflects the combined benefit of the newer processor and platform. It is a measure of generational progress, not an isolated measurement of confidentiality overhead.
Lens 2: performance against a matched general-purpose VM
The second lens holds the processor generation constant. Here, DCas_v6 confidential VMs are compared with Das_v6 general-purpose VMs on the same 4th Generation AMD EPYCTM platform. This is the relevant comparison for understanding how much performance the confidential configuration retained in the tested workloads. The result should not be extended to other processor generations, VM families, or benchmark methods.
AMD-powered Azure VM families covered here
The results in this post are tied to specific Azure VM families. That scope is important: processor generation, VM size, workload placement, and application behavior can all influence the outcome of a performance study.
|
Processor generation |
Confidential VM series |
Largest general-purpose DC (confidential) |
Largest memory-optimized EC (confidential) |
|
3rd Gen AMD EPYCTM 7763v (“Milan”), SEV-SNP |
96 vCPU, 384 GiB |
96 vCPU, 672 GiB | |
|
4th Gen AMD EPYCTM 9004 (“Genoa”), SEV-SNP |
96 vCPU, 384 GiB |
96 vCPU, 672 GiB | |
|
5th Gen AMD EPYCTM 9005 (“Turin”), SEV-SNP |
DCasv7 / DCadsv7, ECasv7 / ECadsv7 (gated preview) |
160 vCPU, 640 GiB |
160 vCPU, 1020 GiB |
Table 1. AMD confidential VM series by processor generation. Specifications from Microsoft Learn VM size documentation. Note: 5th Gen AMD EPYCTM 9005 (Turin), SEV-SNP based VMs discussed here is currently in preview and is not yet generally available.
What changed across AMD confidential VM generations
DCas_v5 to DCas_v6: independent results
In October 2025, Prowess Consulting published Microsoft-commissioned research comparing DCas_v6 confidential VMs on 4th Generation AMD EPYCTM processors with DCas_v5 confidential VMs on the preceding generation. At the benchmark-family level, DCas_v6 delivered 77% higher memory bandwidth in AMD STREAM, 34% higher Redis throughput, and 30% higher SPEC CPU 2017 performance.
Figure 1. Generation-over-generation performance of Azure confidential VMs with SEV-SNP enabled: DCas_v6 on 4th Gen AMD EPYCTM processors versus DCas_v5 on 3rd Gen AMD EPYCTM processors. Results show the geometric mean for each benchmark. Source: Prowess Consulting, Performance and Cost Advantages When Running Latest-Generation VMs, commissioned by Microsoft, October 2025.
Viewed together, these results show broad platform advancement rather than an isolated gain in one test. Organizations moving to the newer AMD confidential VM generation can benefit from stronger performance across compute, memory, and data-service workloads while continuing to use SEV-SNP-based isolation.
Preview: DCas_v6 to DCas_v7
DCas_v7 confidential VMs, built on 5th Generation AMD EPYCTM Turin processors, are being introduced in preview at Microsoft Ignite 2026. To give an early view of what the next generation brings, Microsoft measured DCas_v7 against DCas_v6 at the same vCPU count, with AMD SEV-SNP enabled on both sides. The tests covered Linux and Windows guests at 8 and 32 vCPUs across CPU, cryptography, and web-application workloads.
Across all 22 throughput and score results, DCas_v7 delivered between 12% and 56% higher performance than DCas_v6. Cryptographic hashing showed the largest gains, with SHA-512 throughput up 42% to 56%, while CoreMark and Geekbench 6 multi-core scores rose 27% to 43%. ASP.NET requests per second increased 43% on Linux and 24% on Windows, and SPECrate 2017 integer performance improved 16% to 18% across both operating systems and VM sizes.
Figure 2. Preview of generation-over-generation performance of Azure confidential VMs with SEV-SNP enabled: DCas_v7 on 5th Gen AMD EPYCTM processors versus DCas_v6 on 4th Gen AMD EPYCTM processors at equal vCPU count, for Linux and Windows guests. Results show throughput and score metrics, where higher is better. Source: Microsoft internal testing, October 2026. Preview results; final results may vary.
As with the move from DCas_v5 to DCas_v6, the gains span compute, cryptography, and application workloads rather than a single test. These are preview measurements from Microsoft internal testing rather than an independent study, and they cover a defined set of VM sizes and benchmarks. Customers evaluating DCas_v7 during the preview should validate their own workloads before drawing capacity or sizing conclusions.
How much performance is retained on DCas_v6?
Prowess also evaluated DCas_v6 confidential VMs against Das_v6 general-purpose VMs on the same 4th Generation AMD EPYCTM platform. The confidential configuration delivered 98% of the general-purpose result for AMD STREAM and 92% for both Redis and SPEC CPU 2017. Stated as measured differences, the gaps were 2% for the memory-bandwidth benchmark and 8% for the Redis and CPU benchmark families.
Figure 3. Relative performance of Azure confidential VMs (DCas_v6) compared with same-generation general-purpose VMs (Das_v6) on 4th Gen AMD EPYCTM processors, and the resulting overhead. Results show the geometric mean for each benchmark. Source: Prowess Consulting, Performance and Cost Advantages When Running Latest-Generation VMs, commissioned by Microsoft, October 2025.
These percentages describe the VM configurations and benchmarks included in the study, not a universal overhead for every confidential workload. Application design, VM size, operating system, and workload behavior can change the result. Within the evaluated set, the confidential VM preserved between 92% and 98% of the matched general-purpose VM performance.
What the results mean for platform decisions
AMD confidential VM progress comes from more than a faster processor. Processor improvements raise the compute and memory baseline, while continued evolution in platform architecture, virtualization, and cloud infrastructure shapes the end-to-end experience. The results capture both effects: substantial improvement over the prior confidential generation, continued gains in the DCas_v7 preview, and strong performance retention relative to a same-generation general-purpose VM.
This changes the most useful question for customers. Instead of asking for one universal "confidential computing penalty," ask how the intended workload performs on the current confidential VM against the correct baseline. A defensible answer requires controlled testing at the workload level and a clear distinction between gains from a new generation and differences associated with the confidential configuration.
For AMD-powered Azure Confidential VMs, the data supports three conclusions. First, moving to a newer confidential VM generation can provide meaningful gains over the previous platform, as the Prowess results for DCas_v6 show. Second, early internal results indicate that this progress continues with DCas_v7, which delivered 12% to 56% higher throughput than DCas_v6 in preview testing. Third, the tested DCas_v6 workloads retained a high share of performance when compared with matched Das_v6 general-purpose VMs. Considering those findings together gives customers a clearer basis for evaluating platform progress than a single generalized overhead figure.