Selecting the Right NVIDIA GPU Virtualization Software

NVIDIA GPU virtualization software products are optimized for different classes of workload. Therefore, you should select the right NVIDIA GPU virtualization software product on the basis of the workloads that your users are running.

These tables summarize the differences in features between the various NVIDIA GPU virtualization software products.

Table 13 shows the different classes of graphics workload for which NVIDIA GPU virtualization software products are optimized.

Each NVIDIA GPU virtualization software product is designed for a specific class of workload.

NVIDIA Virtual PC

NVIDIA Virtual PC (vPC) software is designed for knowledge worker VDI workloads to accelerate the following software and peripheral devices:

• Office productivity applications

• Streaming video

• The Windows OS

• Multiple monitors

• High-resolution monitors

• 2D electric design automation (EDA)

NVIDIA RTX Virtual Workstation

NVIDIA RTX Virtual Workstation (RTX vWS) software is designed for professional graphics workloads that benefit from the following NVIDIA RTX vWS features:

• RTX Enterprise platform drivers and ISV certifications

• Support for NVIDIA® CUDA® Toolkit and OpenCL

• Higher resolution displays

• vGPU profiles with larger amounts of frame buffer

NVIDIA RTX vWS accelerates professional design and visualization applications such as:

• Autodesk Revit

• Dassault Systèmes CATIA

• Esri ArcGIS Pro

• Maya

• Petrel

• Solidworks

Using NVIDIA vGPU software to share a GPU among multiple virtual machines improves overall utilization of the GPU. NVIDIA vGPU software uses a best-effort vGPU scheduler by default, which aims to balance performance across vGPUs. In addition to the default scheduler, GPUs based on NVIDIA architectures after Maxwell support equal share and fixed share vGPU schedulers. These scheduling policies allow for more predictable and consistent performance by controlling how GPU resources are allocated to each vGPU.

Performance Allocation for a Shared GPU

In general, when N virtual machines share a GPU, each virtual machine is allocated 1/N of the total performance of the GPU. For example:

• When two virtual machines share a GPU, each virtual machine is allocated approximately 50 percent of the total performance of the GPU.

• When four virtual machines share a GPU, each virtual machine is allocated approximately 25 percent of the total performance of the GPU.

Scheduling Options for GPU Sharing

To accommodate a variety of Quality of Service (QoS) levels for sharing a GPU, NVIDIA vGPU software provides multiple GPU scheduling options. For more information about these GPU scheduling options, refer to Changing Scheduling Behavior for Time-Sliced vGPUs in the NVIDIA Virtual GPU Software User Guide.

Best Effort GPU Scheduler

The default best effort GPU scheduling policy takes advantage of GPU time-slices that other virtual machines have not used. When workloads across virtual machines are not executed at the same time, or are not always GPU bound, the best effort GPU scheduling policy enables a virtual machine’s share of the GPU’s performance to exceed its expected share. Figure 5 is a simplified illustration of how the best effort GPU scheduler works.

Figure 5 - Best Effort GPU Scheduler

Effect of GPU Sharing on Overall Throughput

To measure the effect of GPU sharing on overall throughput, the SPECviewperf 2020 benchmark test was run against a GPU that was allocated to a single VM and then shared among two and four VMs.

Figure 6 shows the results of this test:

• With two virtual machines, throughput is increased by 6%.

• With four virtual machines, throughput is increased by 8%.

Figure 6 - Effect of GPU Sharing on Overall Throughput

This increase in throughput is typical in multi-VM testing scenarios. When scaling from a single VM to multiple VMs, the combined throughput of the VMs should exceed the geomean throughput of the single VM. As additional CPU resources are allocated, throughput improves, peaking at a certain point before stabilizing around 1x throughput.

The server configuration for measuring the effect of GPU sharing on overall throughput is listed in Table 14.