Which software benefits from a GPU VPS
We measured it ourselves: what the card speeds up, by how much, and where it is useless. Bundles from €15.48/mo, 5-hour test
See the measurements
We built two identical machines on the same node — one with a share of an NVIDIA Tesla P4, one without — and ran the same programs on both. The difference in the table below is the contribution of the card: CPU, disk and network were the same on both machines, and the control runs matched to within 0.6%. In short: without a card an Android emulator does not start at all, transcoding runs 3.7–8× faster, 3D and browser graphics give 60 fps against 5 and 33, while Blender rendering is only twice as fast and pure CPU workloads gain nothing at all. Below: what to take for your software and where the card is not needed. A plan plus GPU bundle starts at €15.48/mo, with a 5-hour test period before payment.
Plan plus GPU bundles
The plan and the graphics card go into the cart together
Start
| 4 vCPU · AMD EPYC 7663 |
| 4 GB RAM |
| 40 GB NVMe |
| GPU NVIDIA 1 GB |
| Windows or Linux · RDP |
| Test period 5 hours |
Pro
| 6 vCPU · AMD EPYC 7663 |
| 8 GB RAM |
| 60 GB NVMe |
| GPU NVIDIA 2 GB |
| Windows or Linux · RDP |
| Test period 5 hours |
Max
| 10 vCPU · AMD EPYC 7663 |
| 16 GB RAM |
| 120 GB NVMe |
| GPU NVIDIA 2 GB |
| Windows or Linux · RDP |
| Test period 5 hours |
Measured with and without the card
Two identical machines, the same software versions, median of three runs.
| Software and task | Without a card | With a Tesla P4 share | Difference | |
|---|---|---|---|---|
| LDPlayer 9 — Android boot | does not start, the OpenGL check fails | home screen in 20 s | works / does not work | |
| LDPlayer 9 — installing a 157 MB APK | — | 22 s | — | |
| ffmpeg — H.264 1080p, 52-second clip | 23.1 s, CPU at 80% | 6.2 s, CPU at 2–19% | ×3.7 | |
| ffmpeg — HEVC 1080p | 42.8 s | 5.2 s | ×8.2 | |
| ffmpeg — four 1080p streams at once | not available | 15× real time | — | |
| 3D scene, OpenGL 720p | 5 fps at 79% CPU | 60 fps at 9% | ×12 | |
| 3D scene, Direct3D 11 | does not start (Unknown GPU) | 60 fps | works / does not work | |
| Browser, WebGL over RDP | 33 fps, software renderer | 60 fps, NVIDIA | ×1.8 | |
| Blender 4.2, monster scene | 33.9 samples/min on 6 cores | 65.3 samples/min (OptiX) | ×1.9 | |
| Blender 4.2, classroom scene | 16.7 samples/min | 36.9 samples/min | ×2.2 | |
| Builds, databases, parsing, archiving | the same | the same | ×1.0 — no card needed |
Measured 26–27 October 2026 on our node: AMD EPYC 7663, NVMe, Windows Server 2022, a 2 GB card profile. Control runs (the same software renderer on both machines) differed by 2%, pure CPU encoding by 0.6%.
Android emulators
The main scenario: without a card the emulator fails the OpenGL check. With one, Android is ready in 20 seconds and an APK installs in 22.
Video and streaming
ffmpeg, OBS and anything that speaks NVENC: encoding moves to a dedicated chip on the card and the CPU is freed up.
3D, CAD and the desktop
Viewports, model viewers, WebGL and the Windows interface are drawn by the card — 60 fps instead of 5, and a free CPU.
How to size the configuration
Find your task in the table
If there is a gain, take a bundle with a card. If the row says ×1.0, you do not need one: take a regular plan with more cores instead.
Choose the amount of video memory
1 GB — video encoding, viewers, one or two emulator windows. 2 GB — several windows, a 3D viewport and CUDA compute.
Check it in 5 hours
The machine is handed over before payment: install your own software and verify our numbers on your own workload.
Where the card is not needed — the honest list
Pure CPU work. Project builds, databases, web servers, parsing, archiving — the difference between the two machines was 0.1–0.6%, which means there is none. Buy cores, not a card.
Maximum quality per byte in video. Software x264 scores SSIM 0.99735 against 0.99634 for the hardware encoder at the same bitrate. For archives and master copies the CPU is still better.
Heavy rendering and model training. In Blender a share of the card gives only ×1.9–2.2 over six cores: on 16–40 vCPU the CPU catches up and costs less. For training neural networks 2 GB of video memory is not enough — describe the task and we will size a configuration.
Plans without a graphics cardFrequently asked questions
Didn't find your answer?
Write to us via the contact page.
Two clones of one template on the same node, identical 6 vCPU, 8 GB and NVMe; only the share of the card differed. Each measurement is three runs, and the table shows the median.
There are control rows: pure CPU encoding differed by 0.6%, and with the same software renderer both machines produced 1.20 and 1.23 fps. So the rest of the difference is the card.
The Android emulator: without a card it does not start at all. Video encoding comes next — up to eight times in HEVC, with CPU load dropping from 80% to 2%.
Video encoding, CAD viewers, browser graphics and one or two emulator windows. The amount of video memory does not affect encoder speed at all.
Several emulator windows, a 3D viewport, EEVEE and CUDA compute. In our test four windows used 1.3 GB, so there is headroom.
The Tesla P4 is a 2016 Pascal card with no hardware ray tracing units, so OptiX walks the tree in software. On a many-core plan the CPU catches up without any card.
Not modern ones: 1–2 GB of video memory, no ray tracing and a hard 60 fps ceiling in the profile. Older and lightweight games, plus mobile games in an emulator, are fine.
Yes, the test period is 5 hours: the machine is handed over before payment, you install your software and watch it under real load. If the card did not help, you do not pay for it.