[List of Test Results Obtained in EUC Score Experiments]
Benchmarking Azure Virtual Desktop and the RDP protocol. The test results presented here are a subset of the complete datasets collected during multiple AVD test sequences.
If you want to learn how the EUC Score results were produced then check out EUC Score Test Methodology and EUC Score Toolset.
Testable question: How does an Azure Virtual Desktop D8ads v5 VM compare to the physical lab machine Lancelot without GPU?
Lancelot without GPU
AVD D8ads v5
System Under Test: Physical Lab Machine Lancelot, Windows 10, Intel i7-11700K 16Threads @ 3.6GHz, 64GB RAM, Crucial MX500 SSD 1TB, no GPU.
System Under Test: Azure West Europe, D8ads v5 VM, Windows 11 Enterprise, AMD EPYC 7763 8vCPUs @ 2.4GHz, 32GB RAM, HDD 126GB, no GPU.
Connection: RDP-UDP, 0ms round trip time.
Connection: RDP Shortpath, 15ms round trip time.
Simload and Link | View and Type | Observation | |
---|---|---|---|
![]() |
SL1-ChromeAquariumWebGL | SxS View - Standard | Local RDP connection shows higher frame rate, but also higher network bandwidth. |
![]() |
SL1-ChromeVideoGrid9 | SxS View - Standard | Almost similar user experience. Local RDP connection uses higher network bandwidth while AVD requires more CPU. |
![]() |
SL1-WMPlayer720pMP4 | SxS View - Standard | Different resolutions when video runs with 100% size and both sides freeze when switching to full-screen. |
Result: The overall performance of the physical lab machine Lancelot without GPU and the AVD D8ads v5 VM are quite similar. Only when compute resources are relevant, Lancelot outperforms the AVD session due to higher CPU thread (= core) number and higher clock rate.
Testable question: How does an Azure Virtual Desktop NV4as v4 VM compare to the physical lab machine Lancelot with an NVIDIA M5000 GPU?
Lancelot with M5000 GPU
AVD NV4as v4
System Under Test: Physical Lab Machine Lancelot, Windows 10, Intel i7-11700K 16Threads @ 3.6GHz, 64GB RAM, Crucial MX500 SSD 1TB, NVIDIA Quadro M5000, 8GB VRAM.
System Under Test: Azure North Europe, AVD NV4as v4 VM, Windows 10 Enterprise, AMD EPYC 7V12 (Rome) 4vCPUs @ 2.4GHz, 14GB RAM, HDD 126.45GB, 1/8 AMD Radeon MI25, 2GB VRAM.
Connection: RDP-UDP, 0ms round trip time.
Connection: RDP Shortpath, 28ms round trip time.
Simload and Link | View and Type | Observation | |
---|---|---|---|
![]() |
SL1-DominoOpenGL | SxS View - GPU | The AVD NV4as v4 VM session with app start delay and visible issues, such as rapidly changing frame rates and focus problems. |
![]() |
SL1-FurMarkOpenGL | SxS View - GPU | The AVD NV4as v4 VM session shows lower frame rates and micros studders due to limited GPU resources. |
![]() |
SL1-MSEdgeWaterWebGL | SxS View - GPU | Even though the NV4as v4 VM AVD session uses more CPU and GPU resources, the OpenGL graphics quality is lower. |
Result: The physical lab machine Lancelot with M5000 GPU massively outperforms the AVD NV4as v4 VM. 1/8 of an AMD Radeon MI25 is not sufficient in many use cases to keep up with the reference PC. The limitations of the MI25 graphics driver and the lack of compatibility with some video formats and OpenGL applications is another challenge.
Testable question: How does an Azure Virtual Desktop NV6 VM compare to the physical lab machine Lancelot with M5000 GPU?
Lancelot with M5000 GPU
AVD NV6
System Under Test: Physical Lab Machine Lancelot, Windows 10, Intel i7-11700K 16Threads @ 3.6GHz, 64GB RAM, Crucial MX500 SSD 1TB, NVIDIA Quadro M5000, 8GB VRAM.
System Under Test: Azure West Europe, AVD NV6 VM, Windows 10 Enterprise for Virtual Desktops, Intel Xeon E5-2690 v3 6vCPUs @ 2.6GHz, 56GB RAM, Virtual HD ATA Device 340GB, NVIDIA M60 GPU (1/2 Card), 8GB VRAM.
Connection: RDP-UDP, 0ms round trip time.
Connection: RDP Shortpath, 15ms round trip time.
Simload and Link | View and Type | Observation | |
---|---|---|---|
![]() |
SL1-BSPBlendingDX11 | SxS View - GPU | Almost similar user experience. Higher CPU consumption in the AVD NV6 session, but higher network bandwidth required by local RDP connection. |
![]() |
SL1-FurMarkOpenGL | SxS View - GPU | The AVD NV6 session shows slightly lower frame rates and micros stutters due to limited GPU resources. |
![]() |
SL1-RollercoasterDX9.html | SxS View - GPU | The AVD NV6 session renders only half the frames on the host side, but user experience is almost similar. |
Result: Very similar user experience on both sides, even with demanding use cases. The major reason for this is that the M5000 and M60 graphics cards are using the same NVIDIA GM204 GPU chip with 2,048 shaders. Only the CPU is more powerful on the Lancelot side.
Testable question: What is the performance we can get with an NVIDIA T4-powered AVD VM?
System under test: Azure West Europe, NCasT4 v3 VM, Windows 10 Pro, AMD EPYC 7V12 64-Core Processor 2.4Ghz 8vCPUs, 28GB RAM, HDD 126GB, NVIDIA Tesla T4, 16GB VRAM.
Connection: RDP Shortpath, 15ms round trip time.
Simload and Link | View and Telemetry Type | Observation | |
---|---|---|---|
![]() |
SL1-FurMarkOpenGL | Single View - GPU | Very high frame rate on the host side and on the client side, resulting in outstanding performance and great user experience. |
Result: The AVD NCasT4 v3 VM delivers outstanding performance. Unfortunately, at the time of testing, this VM type lacks availability in many Azure regions and is not officially certified for popular CAD/CAM programs.
Testable question: What can go wrong without a GPU?
System under test: Azure West Europe, D8ads v5 VM, Windows 11 Enterprise, AMD EPYC 7763 8vCPUs @ 2.4GHz, 32GB RAM, HDD 126GB, no GPU.
Connection: RDP-TCP, 23ms round trip time.
Simload and Link | View and Type | Observation | |
---|---|---|---|
![]() |
SL1-CinebenchOpenGL | Single View - Standard | Error: OpenGL application launches, but then terminates without error message due to missing GPU. |
![]() |
SL1-FurMarkOpenGL | Single View - Standard | Error: OpenGL application doesn't launch without error message due to missing GPU. |
![]() |
SL1-WMPlayer720pMP4 | Single View - Standard | Error: 720p MP4 video freezes when user sets it to full-screen. |
Result: These are clearly the wrong use cases for an Azure VM without GPU, typically unnoticed by host-side monitoring systems. Many IT administrators don't know about these effects, that's why screen videos are so important...