20% off every yearly plan this month See the plans →
New machines build in about 60 seconds See plans →
Two new nodes live: Ashburn and Riyadh All locations →
Pay in Bitcoin — no extra fee, 14-day money-back How it works →
Nearest Amsterdam 0.8 ms
Client area See plans

Home/ Dedicated servers/ GPU Server

GPU Server

Whole cards, passed through.

A consumer GPU cannot be split between tenants — no MIG, no vGPU licence. So each machine here hands the card, or all four, to one virtual machine with root on it. Built to order, with you the same working day.

See both machines · built to order, one of each in stock

Two machines, and one question.

The difference is how much machine you get, not what you are allowed to do with it.

Most chosen

1x RTX 4090 · 24 GB

Turbo 2.8 GHz

No terminal limit

8AMD cores, pinned 64 GBDDR5, reserved 240 GB NVMeNVMe Gen4

$539/ mo

2 free Premium Trade Copier FxS Trade CopierFree Copies every order from one terminal to as many as you like, on the same machine. Under 100 ms, nothing leaves the VPS Unlimited slaves, MT4 and MT5 mixed Normally $2.99 a month · included here Trade Insight Trade InsightFree How fast your broker actually fills, measured from inside the machine. Fill time and slippage on every order Which session earns and which one leaks Normally $4.99 a month · included here

$192 in tools, free · over two years

Configure

· built to order, one of each in stock

Measured against the card

What twenty-four gigabytes actually holds.

Rough working figures for 4-bit quantised weights, with room left for context. The bar is one 4090.

7B / 8B ~4.5 GB
13B / 14B ~8 GB
SDXL ~10 GB
LoRA, 7B ~15 GB
30B ~19 GB
70B ~40 GB

RTX 4090 · 24 GB — a 70B model overruns it and spills into system RAM. That is what the 96 GB machine is for.

Before you order

What installs on each, and what does not.

The 96 GB machine has more memory and fewer working tools. This table is the whole trade-off, and it is the thing most likely to catch somebody out.

Software 4090 · CUDA 4× XTX · ROCm
Running models
Ollama The usual way to run a local model Yes Yes
llama.cpp GGUF, quantised weights Yes Yes
vLLM Serving at throughput Yes Partly
text-generation-webui Oobabooga Yes Some backends
Training and tuning
PyTorch The base everything sits on Yes ROCm build
TensorFlow Yes ROCm build
bitsandbytes 4-bit and 8-bit quantisation Yes No
flash-attention Long-context speedup Yes No
xformers Yes No
Unsloth, axolotl Fine-tuning wrappers Yes No
Images, audio, video
ComfyUI SDXL, Flux and the rest Yes Core works
Automatic1111 Yes Extensions vary
Whisper Transcription Yes Yes
Hardware encoding Video transcode NVENC VAAPI
Around it
Docker with GPU Container runtime passthrough Yes Yes
Jupyter, conda Notebooks and environments Yes Yes

Yes installs and runs as documented  ·  Partly works, but expect to patch something  ·  No is CUDA-only, with no ROCm equivalent.

Order to root

Hours, not seconds, and worth the wait.

This one is provisioned by hand rather than by script, because a card has to be bound to a machine and checked before it is yours. Both live in Germany, so there is no city to choose.

01 Order

Tell us what you will run

Not a form and a queue — a short conversation, because the two machines suit different work and the wrong one is an expensive mistake.

0:00
02 Build

We provision the machine

Same working day. The card is passed through, the driver and toolkit go on, and you get root by email.

0:30
03 Verify

Run nvidia-smi yourself

First thing to check: the full VRAM figure, zero used. If it does not read that way, tell us before you install anything.

1:00
04 Run

Install what you need

PyTorch, Ollama, ComfyUI, a container stack. Root means nothing waits on us.

—

The ninety-second build on our other plans does not apply here. One machine, provisioned by hand, and worth the extra hours.

In your panel, not in a ticket queue

Six tools you will use every week.

Buttons you press yourself, included on every plan — this one too. There are more than twenty in the panel; these are the ones that change the day.

The whole card, not a slice

nvidia-smi reports the full VRAM with nothing used. No MIG, no vGPU, no other tenant queued behind your job.

01

CUDA or ROCm, ready

The driver and toolkit are on before handover, because a card machine without them is not a machine you can use.

02

Root and IPMI

Full root from first boot, and the console when the network is the thing that broke. Reinstall the whole machine yourself if you want to.

03

Snapshots before you experiment

A full image on your schedule. Model weights take a while to pull twice, so take one before you change the driver.

04

Move the workload, keep the data

Outgrown the 4090 and want the 96 GB machine? We move the disk rather than make you rebuild.

05

No execution metering

Nothing counts your tokens, your epochs or your hours. The invoice is the same whether the card idles or runs flat out.

06

Twenty more sit alongside them — reboots, graphs, hostname changes, Windows Update control, time zone sync. The cheapest machine has the same panel as the most expensive. See how each one works →

Before you spend anything.

The questions our support desk actually receives, answered plainly.

No. The cards are passed through to your machine and no other tenant can address them. A consumer GPU has no way to be divided — MIG is an A100 and H100 feature, and the vGPU licence does not cover GeForce or Radeon. Anyone advertising "shared 4090 VRAM" is time-slicing it, and your job waits behind theirs.
The 4090 unless you know otherwise. It is cheaper, and the whole Python ML ecosystem is written against CUDA first. Move to the 96 GB machine when you have hit the memory wall and know that you have.
PyTorch, Ollama and llama.cpp all work on ROCm. What does not is anything CUDA-only — bitsandbytes, flash-attention, xformers, most quantisation tooling, and a fair number of projects whose install script assumes NVIDIA. If your workflow is a repository you have not read yet, assume it needs patching.
On the 96 GB machine, yes, resident in VRAM. On the 4090 it loads with layers offloaded to system RAM and runs slowly enough that you will not enjoy it.
The same working day rather than the ninety seconds our other plans quote. Each of these is provisioned when you order it, so tell us what you intend to run and we will have the driver in place.
No. It is against the terms and we will notice. Anything else that fits in the card is fine.
Yes. No minimum term, no setup fee. These are monthly machines like everything else we sell.

Not sure which of the two you need? Tell us what you intend to run and we will say plainly which machine suits it, or whether neither does.