Complete guide, 2026
What is a GPU VPS? Know which slice you’re buying.
Most GPU pages sell you “an NVIDIA GPU” and leave out how much of it is yours. Here’s exactly what our slices are and what they hold.
NVIDIA L4, 24 GB
Forge L4 = one 2 GB slice, 8% of the card’s memory, yours alone. Processing is shared in turns.
NVIDIA L40S, 48 GB
Forge L40S = one 8 GB slice, 17% of the card’s memory, yours alone. Processing is shared in turns.
The short answer
A GPU VPS is a virtual server with access to a graphics card. That access comes in three forms: the whole card, a hardware-isolated MIG slice, or a time-sliced vGPU, where the card’s memory is split into private slices and the processors take turns.
The number that matters most is VRAM, the GPU memory that’s yours. An 8B language model needs about 4 GB for its weights at 4-bit. A LiteKeep Forge L40S gives you 8 GB of dedicated NVIDIA L40S memory, billed by the hour. The calculator below shows what fits.
Here’s the question most GPU server pages skip: how much of the GPU do I actually get?
An NVIDIA L4 has 24 GB of memory. An L40S has 48 GB. A slice gives you a fixed, private part of that memory, and the slice decides what you can run.
So we put it up front. Our Forge L4 is a 2 GB slice of an L4. Our Forge L40S is an 8 GB slice of an L40S. You know exactly what you’re paying for, and you pay for it by the hour.
This guide explains the three kinds of GPU server, how much memory common workloads need, and how to check your slice in five minutes.
In this guide you’ll learn
- Full card, MIG or vGPU: the three things “GPU VPS” can mean
- What a time-sliced slice gives you: private memory, shared processing
- A free VRAM calculator that picks your Forge
- What fits in 2 GB and 8 GB, from the makers’ own figures
- vGPU profiles and CUDA, explained
- How to check your slice with nvidia-smi
Chapter 1
The short answer
“GPU VPS” can mean three different things. Here they are side by side.
| Kind | GPU memory | Processing | Typical use |
|---|---|---|---|
| Full card (passthrough) | All of it, yours | All of it, yours | Training, very large models |
| MIG slice | A fixed part, yours | A fixed part, hardware-isolated | Several isolated jobs on big cards |
| Time-sliced vGPU | A fixed part, yours | Shared, in turns | Inference, speech, image jobs, testing, at a fraction of the price |
Our Forges are time-sliced vGPUs on NVIDIA L4 and L40S data-centre cards. You get your own memory on a data-centre card without paying for the whole card.
Chapter 2
What a vGPU slice really gives you
A time-sliced vGPU comes down to two facts.
- The memory is yours. Once your slice is created, its framebuffer is reserved for you until it’s destroyed. No other customer can use your 2 or 8 GB.
- The processing is shared. Your work and other slices’ work take turns on the card’s engines, so speed can vary with what else is running.
A MIG slice is a hardware-isolated part of a card with its own compute. The L4 and L40S don’t offer MIG, which is why slices on these cards are time-sliced.
| Plan | GPU memory | Share of the card | CPU, RAM, disk | Monthly |
|---|---|---|---|---|
| Forge L4 | 2 GB | 1/12 of an L4 (24 GB) | 8 vCPU, 8 GB, 150 GB NVMe | $40 |
| Forge L40S | 8 GB | 1/6 of an L40S (48 GB) | 32 vCPU, 32 GB, 500 GB NVMe | $400 |
Both are billed by the hour and capped at the monthly price, with 32 TB of traffic, in any of 15 countries, paid in crypto.
Chapter 3
How much VRAM does a model need?
Loading a model’s weights takes its parameter count times the bytes per parameter. For an 8B model that’s 16 GB at 16-bit, 8 GB at 8-bit and 4 GB at 4-bit. The calculator does the same for any size and tells you which Forge holds it.
Weights alone need
4 GB 8B × 0.5 bytes
- Forge L4 sliceLiteKeep Forge, by the hour2 GB: Too big for this slice
- Forge L40S sliceLiteKeep Forge, by the hour8 GB: Weights fit
Weights only. The conversation context (KV cache) needs more: for an 8B model, about 0.125 GB per 1,000 tokens at 16-bit. Billed by the hour, so you can check your stack in the first hour.
The sweet spot: an 8B model at 4-bit is 4 GB of weights, which leaves the Forge L40S’s 8 GB room for a long conversation. Small speech models fit the Forge L4.
Chapter 4
What fits in 2 GB and 8 GB
Here’s what the makers of popular tools say their software needs, against our two slices.
| Workload | Needs | Forge L4 (2 GB) | Forge L40S (8 GB) |
|---|---|---|---|
| Whisper tiny, base (speech to text) | About 1 GB | Fits | Fits |
| Whisper small | About 2 GB | Tight | Fits |
| Whisper medium / turbo | About 5 / 6 GB | No | Fits |
| ComfyUI (image generation) | As low as 4 GB, with offloading | No | Fits |
| Llama 3.1 8B at 4-bit | 4 GB of weights | No | Fits, with room for context |
| Llama 3.1 8B at 8-bit | 8 GB of weights | No | Tight: no room for context |
The short version: the Forge L4 is built for small speech models and for learning GPU setups cheaply. The Forge L40S has room for 8B language models at 4-bit, Whisper medium and image generation.
Chapter 5
vGPU profiles and CUDA
A vGPU slice comes in a profile series, and the series decides which software stack it’s built for, including CUDA, the toolkit most AI frameworks use.
| Series | Made for | CUDA |
|---|---|---|
| Q | Workstations and compute | Yes |
| C | Compute and AI | Yes, from 4 GB on the L4 |
| B | Business desktops | No |
| A | Virtual applications | No |
On the L4, a 2 GB slice can be Q, B or A. On the L40S, 8 GB exists as both Q and C.
Chapter 6
Check what you got
On any GPU server, check these before you rely on it:
- Is the driver loaded?How to check: Run
nvidia-smi. If it says “command not found” or can’t talk to the driver, the guest driver isn’t installed or loaded. - How much memory is yours?How to check: The memory column in
nvidia-smishows your slice’s total, for example about 2 GB on a Forge L4. - Is it a vGPU?How to check:
nvidia-smi -qlists the virtualization mode. - Does CUDA work?How to check: Install the CUDA toolkit from NVIDIA’s Linux installation guide and run a small sample, or check that your framework sees the GPU.
- Under real load?How to check: Run your actual job for an hour and watch nvidia-smi. Time-sliced processing means speed can vary.
Chapter 7
NVIDIA L4 vs L40S
Both are data-centre cards from NVIDIA’s Ada Lovelace generation. Here are the full cards; each Forge is a slice of one.
| NVIDIA L4 | NVIDIA L40S | |
|---|---|---|
| GPU memory | 24 GB | 48 GB GDDR6 with ECC |
| Memory bandwidth | 300 GB/s | 864 GB/s |
| Power | 72 W | 350 W |
| Your Forge slice | 2 GB (Forge L4) | 8 GB (Forge L40S) |
The L4 is a small, efficient card built for inference and video. The L40S is a big all-rounder with almost three times the memory bandwidth. At slice sizes, memory matters most: pick the slice that holds your model.
Chapter 8
Is a GPU slice right for you?
Slices are cheap per hour because you share a big card. Here’s how to pick.
- Forge L4: small speech models, learning and testing GPU setups, short jobs billed by the hour.
- Forge L40S: 8B language models at 4-bit, Whisper medium, image generation, and anything that fits in 8 GB.
- Not a GPU job at all? AI agents that call a hosted model need no GPU. A CPU Lite is enough, from a few dollars a month.
Now it’s your move
“GPU VPS” hides a lot. The memory that’s yours decides what runs, and whether processing is shared decides how fast.
Our answers are on the label: 2 GB or 8 GB of dedicated NVIDIA memory, time-sliced, billed by the hour, in 15 countries, paid in crypto. Deploy a Forge and test your workload today.