Nvidia tried so hard to stop this - GPU Sharing with Virtual Machines

Overview

The video demonstrates how to turn one Windows gaming PC into multiple virtual gaming machines (VMs), allowing several people to share one GPU. The setup uses Windows Hyper-V and GPU paravirtualization, with the community script Easy-GPU-P automating much of the VM configuration. Each VM runs games on the shared hardware, while Parsec streams its display to another device.

How the Setup Works

• NVIDIA has historically limited GPU virtualization on consumer GeForce cards, reserving supported features for higher-priced workstation and data-center products.
• Windows paravirtualization lets VMs access part of the host GPU through Hyper-V. Easy-GPU-P helps create and configure VMs, install drivers, and set up Parsec.
• VMs can share GPU processing power and memory, but they cannot use the GPU’s physical display outputs. Parsec streams each VM’s display instead.
• Small HDMI or DisplayPort “dummy plugs” can make a VM think a monitor is connected, allowing higher resolutions to be selected.
• GPU video memory must be divided evenly among the VMs, while system RAM can be allocated more flexibly.

Demonstrations and Performance

The main test system used a Ryzen 9 5950X, an RTX 3080 Ti, and 32 GB of RAM. The team tested two and then four simultaneous gaming sessions.

• With two CS:GO VMs, the combined frame rate was reported at roughly 350–400 fps. A single bare-metal run averaged about 460 fps, while a VM using half the GPU managed around 175 fps. This shows that performance does not scale perfectly with the GPU share.
• In a four-player test, the sessions produced around 300 fps combined, but the person playing directly on the host had a more responsive experience than the streamed players.
• Anno 1800 ran at roughly 40–50 fps under the demonstrated conditions, with a reported 99th-percentile figure of 25 fps.
• A less expensive RTX 2070 Super was also described as surprisingly usable for lighter games such as Rocket League.

The results show that GPU sharing can work, but allocating a quarter of a GPU to a VM does not guarantee a quarter of its single-machine performance. Video encoding, CPU allocation, game demands, and streaming all affect the results.

Setup Guide

1. Use Windows 10 or 11 Pro or Enterprise; the presenters recommend Windows 11.
2. Download Easy-GPU-P from its GitHub page and obtain the latest GPU drivers.
3. Enable CPU virtualization in BIOS/UEFI—typically SVM on AMD or Intel VT-x on Intel.
4. Enable Hyper-V and its required Windows features, then restart.
5. Configure and run the Easy-GPU-P scripts to create VMs. Give each VM a unique name if creating several.
6. Start the VMs in Hyper-V Manager, install games, and configure Parsec. Use dummy display plugs if higher virtual-display resolutions are needed.

The video notes that the script and setup steps may change, so it directs viewers to the project’s GitHub instructions and the Linus Tech Tips forum for current guidance.

Limitations and Costs

• Consumer NVIDIA cards limit concurrent encoding streams to three, making the stated practical maximum four players total: one local and three remote.
• Parsec adds latency—about a frame on a local network and a few frames remotely—and streamed video loses image quality compared with a direct display connection.
• A reliable wired network is recommended. For hosting two or more VMs over the internet, Parsec suggests more than 50 Mbps upload speed.
• Windows licensing is required for each VM.
• Sharing one PC can reduce the need for multiple cases, power supplies, and SSDs, but the host still needs capable components. The presenters warn that storage, power delivery, motherboard, and cooling demands rise when multiple users load the system at once.

Overall, the video presents GPU sharing as a novel, potentially cost-saving way to create a small private cloud-gaming setup, while emphasizing its setup constraints, added latency, and uneven performance.

Main Speakers and Sources

• Linus Sebastian — presenter and primary narrator.
• Colin — co-presenter and the person who set up the test.
• David, Jonathan, and Jeff — participants in the gaming demonstrations.
• Easy-GPU-P — the community script discussed in the tutorial, attributed in the video to a Parsec developer.
• NZXT — sponsor of the video and provider of the test PC.