A cheap GPU VPS for everyday GPU workloads

What is inside the VPS

A KVM virtual machine running Debian 13 (kernel 7.3) with the Intel UHD 770 integrated GPU exposed inside the VM: no emulation, the standard Linux drivers talk straight to the hardware.

• GPU: Intel UHD 770, Xe-LP (Gen12) architecture, 32 EUs (256 FP32 lanes), up to 1.45 GHz, ≈ 0.74 TFLOPS FP32 peak. No dedicated VRAM: the GPU uses the VM’s RAM.
• Devices: /dev/dri/card0 and /dev/dri/renderD128, usable from Docker and from k3s pods.
• Software stack: Mesa iris (OpenGL 4.6) and ANV (Vulkan), intel-media-driver iHD (VA-API), compute-runtime NEO (OpenCL 3.0, Level Zero), OpenVINO GPU plugin.
• Media engine (QuickSync): decode H.264, HEVC 8/10-bit, VP9, AV1; encode H.264, HEVC 8/10-bit, VP9. No AV1 encoding.
• System: root SSH, Docker preinstalled, dedicated public IPv6, outbound IPv4 (dedicated IPv4 +€5/month).

Plans

The GPU is time-shared between a small number of VMs. Your plan sets your share of GPU time (Perf 1, 2 or 3). For a card of your own, pick the GTX 1060.

* Shared GPU: several customers may use the same physical GPU. The split is fixed by SR-IOV (hardware GPU virtualization): each VM gets its own share of the GPU, sized by the plan you choose (Perf 1, 2 or 3).

What a GPU VPS is good for

The workloads where a GPU makes the difference are not always the ones that need tens of TFLOPS:

• Video transcoding (FFmpeg, Jellyfin, Plex): the fixed-function encoder does the work, not the shader cores. 1080p and 4K HEVC through VA-API, with very little CPU.
• Browsers with a GPU (Chrome, Playwright, Selenium, AI agents): in our measurements on a Perf 2 plan, a WebGL benchmark runs at 60 fps on 0.3 vCPU with the GPU, versus 9.8 fps on 3.4 vCPU with software rendering (SwiftShader).
• Small AI models: vision, OCR, embeddings with OpenVINO; 3–8B-parameter LLMs quantized to Q4 (Gemma, Phi, Qwen) with llama.cpp on Vulkan or Ollama. For asynchronous jobs (summaries, extraction), not for high-traffic chat.
• Headless rendering (Blender EEVEE, OpenGL/Vulkan), OpenCL compute (Julia, PyOpenCL), graphical CI.

Check the GPU in 10 seconds

Once logged in over SSH:

gpu-check            # preinstalled: GPU, VAAPI, OpenCL, Vulkan in one go
ls -l /dev/dri        # card0 + renderD128
vainfo --display drm  # iHD driver, H.264/HEVC/VP9/AV1 profiles
clinfo -l             # Intel(R) OpenCL Graphics (NEO)
vulkaninfo --summary  # Mesa ANV, Intel(R) UHD Graphics 770

Flat monthly price vs hourly billing

GPU clouds show hourly prices. A service that runs all the time (media server, browser fleet, CI runner) is billed 720 hours a month: at $0.60/h that is ≈ $432/month; at $1.00/h, ≈ $720/month. The GPU VPS stays at €14.99/month, with no setup fee.

Detailed comparison: Hetzner, OVHcloud, Vultr, Vast.ai →

When you need something bigger

To be clear about the limits: no CUDA on the iGPU, no dedicated VRAM, ≈ 0.74 TFLOPS. For training, 30B+ parameter LLMs or high token throughput you need 24–96 GB of VRAM: an L4, an L40S or an RTX PRO from a large cloud. For CUDA and NVENC on a budget, our GTX 1060 (€49.99/month) is often enough.