Video summary

Nvidia Tesla P100 vGPU Cloud Gaming Performance!

Main summary

Key takeaways

Technology

Summary of Technical Content (NVIDIA Tesla P100 vGPU for Cloud Gaming)

Target Use Case

This review evaluates the NVIDIA Tesla P100 as a cost-effective enterprise GPU for cloud gaming, particularly in setups where the GPU runs in a server and streams video to clients (PCs/handhelds) using tools such as Sunshine and Moonlight.

Why Enterprise GPUs Can Still Work for Gaming

The creator argues that many enterprise GPUs map closely to consumer GPU designs because they often share the same underlying GPU dies and VRAM configurations. Differences are mostly server-focused:

  • Cooling: Often passively cooled, relying on server chassis airflow.
  • Power input: Uses EPS 12V connectors (server CPU power rails) rather than standard PCIe 8-pin, reducing cable clutter and matching typical server power delivery.

Tesla P100 Hardware Highlights

Purchase Context / Price

  • The creator monitored the secondhand market (eBay).
  • Bought the card for ~$165 after it fell below $200.

GPU + Memory Architecture

  • GPU die: GP100
  • Memory: 16GB HBM2
  • Key spec: HBM2 provides an extremely wide ~4096-bit memory bus, compared with typical 384-bit GDDR setups.

Launch Cost vs. Today

  • Originally expensive: about ~$5,700 at launch (June 2016).
  • Today it’s considered usable due to significantly reduced used pricing.

vGPU Setup Approach (Proxmox + NVIDIA Licensing/Driver Behavior)

Setup References

  • Mentions a prior tutorial for manually installing NVIDIA vGPU / NVIDIA GRID.
  • References an installer script (linked in the description) that:
    • Configures Proxmox
    • Installs required drivers
    • Sets up vGPU
    • Includes a vGPU unlock mod for cards that aren’t officially supported (Note: the P100 is described as officially supported, so the official driver route should work.)

Server Platform Used for Testing

  • CPU: AMD EPYC 7742 (64 cores)
  • RAM: 256GB DDR4 registered ECC
  • Hypervisor: Proxmox 8.1.4

vGPU Profile Configuration in Proxmox (GUI / Hardware Passthrough)

The configuration uses hardware passthrough for mdev (mediated devices):

  1. Add the PCI device
  2. Select NVIDIA P100 from the dropdown
  3. Choose the mdev type / vGPU profile from available options

This is positioned as an improvement over older workflows that required SSH-editing VM config files and manually defining profiles.

vGPU Partitioning and Performance Behavior

Profile Types

  • There are three main profile types
  • The creator recommends Q-type profiles for most use cases.

Memory Slicing

With 16GB, the P100 can be split into:

  • two 8GB partitions, or
  • smaller slices as needed.

Dynamic GPU Power Sharing (Multi-VM Important)

When multiple VMs exist, GPU utilization is dynamically balanced. For example:

  • One VM may reach up to ~100% compute power when another VM isn’t actively running 3D.
  • When another VM starts gaming, performance tends to equalize.

Streaming and Driver Constraints Used in Testing

Streaming Stack Concept

The broader service concept includes:

  • Linux Mint + Proxmox + Sunshine + Moonlight

Windows VM Driver Choice for Testing

Inside Windows VMs, the test uses NVIDIA’s GRID client drivers. These virtual drivers provide a virtual display required for remote streaming, but impose limitations:

  • Maximum refresh rate: 60 Hz
  • Max FPS cap: 60 FPS (for rasterization)

Implications

  • Benchmark scores are expected to be lower than bare-metal results due to the 60 FPS environment cap.

Mentioned Workarounds (Not Used)

  • Third-party virtual display drivers
  • Spoofing the Nvidia PCI ID passed to the VM

Benchmark Methodology (Key Details)

Test Scenario

  • Running two Windows 11 VMs simultaneously.

Workload Pattern

  • Baseline: Single VM running a game (via Sunshine streaming)
  • Scaling / Stress: Second VM runs Unigine Heaven Benchmark in a loop (representing a gaming-like workload)

Example Settings / Resolution

  • Heaven (VM2): 1080p High, normal tessellation, 2x anti-aliasing

Goal

Compare:

  • The absolute performance ceiling within the 60 FPS constraint
  • Scalability when both VMs are active

Performance Results (Single vs Dual VM Scaling)

3DMark Fire Strike

  • Solo (VM1 only):
    • Overall: 12451
    • Graphics: 13799
  • Dual VM:
    • Overall: 11505
    • Graphics: 12200
  • Observed drop: ~10%, described as strong scaling

3DMark Time Spy

  • Approximately 30% drop with both VMs running
  • Still framed as comparable to around a Radeon RX 7600 class result (per creator’s context)
  • Note: scores are affected by the 60 FPS limitation even during 3DMark testing

Game Examples (Single VM vs With Heaven on VM2)

Crisis (Crysis)

  • Solo: ~56 FPS avg, lows into the teens
  • Dual VM: minimal change; averages not framed as meaningfully improved or worsened

Red Dead Redemption 2

  • Solo: ~55 FPS avg, 18 low
  • Dual: avg drops to ~34 FPS
  • Still considered “playable”; lowering settings can improve FPS

Cyberpunk 2077 (No Ray Tracing; uses FSR 2)

  • Solo: 59 FPS avg, 19 low at 1080p high + FSR 2 (described as very impressive)

  • Dual (Heaven on VM2): ~40 FPS avg, low ~16

  • Creator suggests tuning could make it more than playable for multi-instance setups

Other Titles Mentioned (Multi-Client Scaling / “Ultra-ish”)

  • Doom? (text unclear: “duom maternal”)
    • Solo: ~60 FPS avg, low 31
    • Dual: keeps 60 FPS avg, low 28
  • Tiny Tina’s Wonderlands
    • Solo: ~55 FPS avg, low 27
    • Dual: ~50 FPS avg, low 29
  • Reccfest
    • Solo: 60 FPS avg, low 49
    • Dual: avg ~60, low 36

General Conclusion from Results

The P100 is presented as capable of supporting multiple 1080p gaming sessions on one GPU while preserving strong usability—especially given the test workloads and the 60 FPS streaming/driver cap.

Final Product / Review Takeaway

Usefulness for Cloud Gaming

At about ~$165, the Tesla P100 is positioned as a practical way to build a server GPU that can stream two 1080p games to client devices (including lower-end clients) without each client needing its own dedicated GPU.

Future Content

The creator plans a tier list / ranking comparing the P100 against other previously tested cards for vGPU performance and value at today’s used prices.

Main Speakers / Sources

  • Speaker: Jeff (host of Craft Computing)
  • Referenced tools / sources: NVIDIA GRID/vGPU, Proxmox, Sunshine, Moonlight, Linux Mint, an installer script linked by the creator, and Maximum Settings (sponsorship and cloud gaming offering)

Original video