Blog Post

Azure Confidential Computing Blog
6 MIN READ

Closing the Azure Confidential VM Performance Gap with Intel® TDX

Mihir22_Patel's avatar
Mihir22_Patel
Icon for Microsoft rankMicrosoft
Nov 18, 2026

The question every customer asks

When organizations evaluate confidential computing, the conversation usually starts with security: how memory is protected, how attestation works, and who controls the root of trust. But it rarely ends there. The next question is almost always performance.

Customers want to know what stronger confidentiality means for workload and application throughput, latency, and infrastructure efficiency. They want to understand whether protecting data in use creates a meaningful performance tradeoff and whether that tradeoff is changing over time.

Recent Intel performance publications, together with internal Azure engineering data on local NVMe and Accelerated Networking, provide a more useful way to frame the discussion. Rather than asking for one universal “confidential computing overhead” number, we can examine how workload behavior and platform capabilities affect end-to-end performance on Confidential VMs with Intel TDX.

  1. How close is a Confidential VM with Intel TDX to a matched general-purpose VM?
  2. How do local NVMe and Accelerated Networking improve end-to-end I/O performance?

Those questions are related, but they are not interchangeable. Keeping them separate is essential to interpreting the data correctly.

The right comparison: confidential versus general purpose

A same-generation comparison estimates the end-to-end performance difference associated with choosing the confidential VM configuration. The Intel data discussed below compares DCeds_v6 confidential VMs with Dds_v6 general-purpose VMs on the same 5th Generation Intel Xeon Scalable processor platform. This comparison is the appropriate basis for discussing the measured performance retained by the Confidential VM with Intel TDX in the tested workloads.

The Azure confidential VM families in scope

This post focuses on the Azure Intel TDX VM families used in the published performance comparison. The catalog context matters because processor generation, network and storage configuration, VM size, and workload placement all influence the measured result.

Processor Generation

Confidential VM series

Largest General Purpose DC (Confidential)

Largest memory optimized EC (Confidential)

5th Gen Intel® Xeon Platinum 8573C (Emerald Rapids), TDX

DCesv6 / DCedsv6, ECesv6 / ECedsv6

128 vCPU, 512 GiB

64 vCPU, 512 GiB

Table 1. Confidential VM with Intel TDX series. Specifications from Microsoft Learn VM size documentation.

The Intel picture: same-generation performance

Intel’s published comparison uses Azure DCeds_v6 confidential VMs and Dds_v6 general-purpose VMs on 5th Generation Intel Xeon Scalable processors. The evaluation matched 16-vCPU, 64-GB configurations with local NVMe backing, Windows Server 2025 and Ubuntu 24.04, and three runs per test.

The confidential VM retained 97% to 99% of general-purpose performance across the reported SPEC CPU 2017 rate-suite summaries. The NGINX-WordPress workload retained 97%, while PostgreSQL driven by HammerDB retained 86%.

 

Figure 1. Relative performance of Azure DCeds_v6 confidential VMs compared with same-generation Dds_v6 general-purpose VMs on 5th Gen Intel Xeon Scalable processors. SPEC CPU 2017 results are geometric means; web and database results are throughput ratios. Source: Intel Corporation, document 926026, August 2026. Chart created from the reported benchmark results.

The range reinforces why workload-level characterization matters. The reported CPU rate suites and web stack were close to the matched baseline, while the PostgreSQL-HammerDB workload showed a larger difference. The published chart establishes the measured difference, but customers should avoid treating one database result as representative of every transactional application or storage configuration. Intel plans to reevaluate PostgreSQL-HammerDB performance once Accelerated Networking through MANA is available for this Intel TDX configuration, enabling the workload to be assessed with the optimized network path.

Optimizing I/O performance for Confidential VMs with Intel TDX

The published Intel results establish a compute and multi-tier application performance baseline for the TDX configuration, but they do not capture the broader impact of platform co-optimization. Two capabilities available on the Intel confidential VM platform illustrate how coordinated hardware and software enablement can improve end-to-end workload performance: local NVMe for storage and Azure Accelerated Networking through Microsoft Azure Network Adapter (MANA) for networking, which is currently available in preview for the Confidential VM with Intel TDX configuration discussed in this post.

Local NVMe: a shorter storage pathlength

In the earlier virtualized-SCSI design, a request traverses the guest SCSI stack and VMBus into OpenHCL, where it is translated before reaching the underlying NVMe device. The local-NVMe path presents an NVMe interface to the guest through vPCI, removing the virtual-SCSI protocol conversion and VMBus transport from the I/O path.

In the measured storage data, local NVMe delivered nearly 5 times the IOPs of the single-thread virtualized-SCSI baseline and reduced average I/O latency by approximately 16%.

Figure 2. Local NVMe throughput and latency improvement over virtualized SCSI. Linux confidential VM, random reads, 8 KB block size, queue depth 1, single thread. Source: Microsoft Azure internal engineering data.

At higher concurrency, the same internal engineering data shows near-parity in local-NVMe performance between matched general-purpose and confidential VM configurations. At both 32 and 64 threads, the confidential VM delivers comparable IOPS and average I/O latency, demonstrating general-purpose-class local-NVMe performance for the tested workload. 

Figure 3. Local NVMe IOPS and average-latency comparison at 32 and 64 threads. Windows confidential VM and matched general-purpose VM, 8 KB random reads, queue depth 16. Source: Microsoft Azure internal engineering data.

These results demonstrate how new platform capabilities enabled through hardware and software co-optimization can substantially improve confidential VM performance. With local NVMe, the tested confidential VM achieved throughput and average latency comparable to the matched general-purpose VM at 32 and 64 threads.

Accelerated Networking: removing the synthetic throughput ceiling

With the synthetic NIC path, packets traverse the guest-facing synthetic interface and host networking path. Accelerated Networking exposes a MANA SR-IOV virtual function to the guest, bypassing the host virtual switch for the fast datapath while network policy remains enforced in adapter hardware.

The measured networking data shows a throughput benefit from MANA at low thread counts: approximately 3.20 times the synthetic-NIC throughput at one thread, 2.00 times at two threads, and 1.22 times at four threads. Average latency was approximately 24% lower than the synthetic path. 

Figure 4. Accelerated Networking throughput and average-latency improvement over the synthetic NIC. Source: Microsoft Azure internal engineering data. Note: Accelerate Networking with MANA for the Intel TDX configuration discussed here is currently in preview and is not yet generally available.

At higher concurrency, MANA enables confidential VM network throughput to match the general-purpose baseline. In the 128 KB NTTTCP sweep at 32 threads, the confidential VM with MANA achieved 100% of the matched general-purpose VM’s throughput, compared with approximately 74.1% for the confidential VM using a synthetic NIC. In this tested configuration, Accelerated Networking increased throughput by 35% over the synthetic path and delivered throughput parity with the matched general-purpose VM. 

Figure 5. NTTTCP throughput at 32 threads with a 128 KB buffer: matched general-purpose MANA, confidential synthetic NIC, and confidential MANA configurations. Source: Microsoft Azure internal engineering data.

Bandwidth parity does not mean every remaining performance difference has disappeared. The internal analysis shows a larger throughput gap at low thread counts, along with CPU-efficiency and latency differences even when throughput converges at higher concurrency. MANA restores bandwidth scalability, while efficiency and latency remain dependent on the workload and concurrency level.

A broader definition of platform progress

The evolution of confidential VM performance is not only a processor story. Processor capabilities establish the compute and memory baseline, while platform capabilities enabled through Intel and Azure hardware and software co-optimization improve overall performance, efficiency, and scalability. Local NVMe and Accelerated Networking illustrate this broader progress by reducing storage-path overhead and removing the synthetic-network throughput ceiling. Emerging platform innovations such as Azure Boost Confidential Device and standards-based technologies like TDISP further advance this model by offloading infrastructure functions and enabling more efficient, secure device interactions within confidential computing environments. Together, these advances preserve the confidential trust boundary while bringing more general-purpose-class performance capabilities into the Confidential VM with Intel TDX environment. The practical question is therefore not, “What is the confidential computing penalty?” It is, “How much performance does my workload retain on the current Confidential VM with Intel TDX configuration compared with the right baseline?” Answering that question requires workload-level measurement, a controlled baseline, and a clear separation between compute behavior, platform-capability gains, and confidentiality-specific cost. For customers evaluating Confidential VMs with Intel TDX, the data points to a clear approach: compare against a matched same-generation baseline, characterize the actual workload, and include the I/O capabilities that shape end-to-end performance. The result is a more accurate view of the security-performance tradeoff than any single universal overhead number can provide.

Sources and further reading

Updated Nov 13, 2026
Version 1.0
No CommentsBe the first to comment