Bare-metal NVIDIA GPU servers, configured your way
KW Servers GPU dedicated servers give you a whole physical machine with NVIDIA accelerators for LLM inference, model training, 3D rendering, video transcoding and virtual desktops. Pick a ready configuration or build your own, choose a location, and we deploy in 24 to 48 hours.
A GPU dedicated server pairs one or more graphics processing units with a server-grade CPU, plenty of RAM and fast NVMe or SSD storage. Because the machine is bare metal, nothing is shared and nothing is virtualized: the full GPU, CPU and RAM work for your workload, and performance stays consistent from the first day to the last.
You stay in control of the software. Choose the operating system, pin the NVIDIA driver and CUDA version your code was tested against, and install a custom image through IP KVM if you need one. Whether you are fine-tuning a language model, serving inference to customers, rendering frames or running simulations, you get hardware that is yours alone.
Uptime and network performance are backed by a written SLA, so long training runs, render queues and live inference endpoints stay online.
Direct connections to major providers and exchanges keep latency low for your users and speed up dataset and checkpoint transfers.
H100 NVL, A100, A30, L40S, A40, L4 and T4, plus AMD accelerators on selected builds, so you can match the card to the job instead of overpaying.
Fast local storage keeps datasets, model checkpoints and render caches from starving the GPU.
Start with one card and move to a dual-GPU build when a model no longer fits in one card's memory or training needs to go faster.
Choose an operating system, pin your own CUDA and driver versions, or install a custom image through IP KVM. Nothing changes underneath you.
A fixed monthly server is usually cheaper than per-hour cloud GPUs once a workload runs most of the month.
Training and fine-tuning need memory first, then bandwidth. Choose H100 NVL or A100, and a dual-GPU build when one card is not enough. Fast NVMe storage keeps data loaders from becoming the bottleneck.
Match VRAM to model size using the table below. L40S, A30 and H100 NVL cover most self-hosted LLM and RAG deployments, and quantized models fit smaller cards at a modest cost in quality.
A40, L40S and L4 combine large frame buffers with hardware video encoders, which suits 3D rendering, compositing, live streaming and transcoding pipelines.
T4 and A40 power virtual desktops and remote workstations for design, engineering and CAD teams that need smooth graphics from anywhere.
Accelerate simulations, analytics and notebooks with A100 or A30 plus fast local NVMe storage. Parallel workloads in physics, chemistry and engineering benefit the most.
L4, A40 and L40S handle cloud game streaming, headless game clients and build-and-test rendering for game and VR developers.
Choose the card by the memory your workload needs first, then by the kind of work it does. Every model below can be deployed on a dedicated server, and availability varies by location.
The H100 NVL is a Hopper-generation accelerator built for large language models, computer vision, retrieval-augmented generation and conversational AI. Its 94 GB of memory lets bigger models stay on a single card, which simplifies deployment and reduces the need to split a model across GPUs.
The L4 is a low-power card with 50% more memory than the T4. It is a cost-aware choice for video streaming, encoding, image processing and small inference models.
The L40S is a flexible Ada-generation card that handles AI inference, fine-tuning smaller models and graphics on the same server. Its 48 GB leaves room for mid-sized language models and image generation workloads, and it also renders and encodes video well.
The T4 is a compact, efficient entry point for virtual desktops, light inference and mixed compute and graphics work, with low power draw and broad compatibility.
These Ampere accelerators remain a proven choice for deep learning, scientific computing and data analytics. The A30 suits inference and mid-sized training jobs, while the A100 gives extra memory and bandwidth for heavier models. Both are available on selected double-width GPU servers.
The A40 pairs a large frame buffer with strong graphics output, which makes it a good fit for virtual workstations, 3D rendering, VR projects and video pipelines. It is available on selected double-width servers.
| GPU | Memory | Best for |
|---|---|---|
| NVIDIA H100 NVL | 94 GB HBM3 | Large language models, training, RAG |
| NVIDIA A100 / A30 | A100 40/80 GB, A30 24 GB | Deep learning, HPC, analytics |
| NVIDIA L40S | 48 GB GDDR6 | Generative AI, smaller LLMs, rendering |
| NVIDIA A40 | 48 GB GDDR6 | VDI, 3D rendering, transcoding |
| NVIDIA L4 | 24 GB GDDR6 | Video, image processing, light inference |
| NVIDIA Tesla T4 | 16 GB GDDR6 | VDI, entry-level inference |
GPU memory decides whether a model loads at all, while cores and clocks decide how fast it runs. A card that is two gigabytes short does not run slowly; it fails to load. Settle the memory number first.
Covers the context cache and runtime. The cache grows with every token of context.
| Model size | 16-bit | 8-bit | 4-bit |
|---|---|---|---|
| 7B parameters | 14 GB | 7 GB | 3.5 GB |
| 13B parameters | 26 GB | 13 GB | 6.5 GB |
| 34B parameters | 68 GB | 34 GB | 17 GB |
| 70B parameters | 140 GB | 70 GB | 35 GB |
7B or 13B at 8-bit runs comfortably on a card such as the L4 or A30.
34B at 4-bit fits a 24 GB card with limited context. A 48 GB L40S or A40 gives more room.
70B at 4-bit needs about 35 GB for weights, so a 48 GB card is the sensible minimum.
70B at 16-bit needs multiple GPUs or a dual-GPU build.
If the number lands between cards, choose the larger one or a dual-GPU build. Quantizing to 4-bit lets a model fit a smaller card at some cost in output quality. For hands-on setup, follow our guides to self-hosting an LLM on a dedicated server and deploying Ollama and Open WebUI on multiple GPUs.
Both options have a place. The right one depends on how steadily your workload runs.
| Dedicated GPU server | Cloud GPU instance | |
|---|---|---|
| Billing | Fixed monthly cost | Usually per hour |
| GPU access | The whole physical card | Often shared or sliced |
| Best for | Training, inference and rendering that run most of the month | Short experiments and bursts |
| Software control | Choose OS, driver and CUDA version | Often limited to provider images |
| Performance | Consistent, no noisy neighbours | Can vary with neighbouring load |
If your GPU sits busy for most of the month, a dedicated server is usually the more predictable and cost-effective choice.
Use the configurator to pick a ready configuration, or customize CPU, RAM and storage.
Most GPU servers are deployed in 24 to 48 hours, depending on the card, configuration and region.
Log in with full control, choose your operating system, and install drivers, CUDA and your frameworks.
Move to a larger card or a dual-GPU build as your models and traffic grow.
The server is yours to configure. Install NVIDIA drivers and CUDA, then PyTorch, TensorFlow, Ollama, JupyterLab or your own containers on Ubuntu, Debian, AlmaLinux or Windows Server. Pin the versions your code was tested against so nothing changes underneath you, and use IP KVM if you need to install a custom image or recover a misconfigured driver.
Planning for newer hardware? Read NVIDIA Rubin R200 vs Blackwell B200, the B200 setup guide for Ubuntu 24.04.
Deploy near your users or data to cut latency and meet local data requirements.
Confirm GPU availability for your chosen city in the configurator.
Answers on sizing, deployment and choosing between GPU models.
A GPU dedicated server is a bare-metal server with one or more NVIDIA or AMD GPUs next to the CPU, RAM and storage. It is rented to one customer, so nothing is shared or virtualized. That suits parallel workloads such as AI training, inference, rendering and simulation.
Multiply the parameter count by the bytes per parameter: 2 bytes at 16-bit, 1 at 8-bit and about 0.5 at 4-bit. Then add 25 to 50% for the context cache and runtime. A 7B model fits a 16 to 24 GB card. A 70B model at 4-bit needs roughly 35 GB for weights, so a 48 GB card is the sensible minimum.
For inference, L40S, A30 and L4 give good throughput per dollar when the model fits in memory. For training and very large models, H100 NVL and A100 offer more memory and compute. Tell us your model and context length and we will recommend a configuration.
For workloads that run most of the month, usually yes: you get a fixed monthly cost, the full card, and control of your driver and CUDA versions. Cloud instances fit short experiments and bursts.
A regular dedicated server relies on the CPU alone. A GPU server adds one or more accelerators that run thousands of operations in parallel, which makes training, inference, rendering and simulation much faster than on CPU-only hardware.
Yes. Single and dual-GPU configurations are available on supported servers. A second card helps when a model exceeds one card's VRAM or when you want faster training.
NVIDIA H100 NVL, A100, A30, L40S, A40, L4 and T4, with AMD accelerators on selected builds. Availability depends on location, and the configurator above shows what can be deployed.
You choose. Install the driver and CUDA release your code was tested against, and keep them pinned. If you need a specific version in place before you start, tell us when you order.
Usually 24 to 48 hours, depending on the GPU, configuration and region.
Ubuntu 22.04 and 24.04, Debian 11 and 12, AlmaLinux 8, 9 and 10, CentOS 7.9, 8 and 9, Windows Server 2019 and 2022, and VMware ESXi 7 and 8. You can also install any custom OS through IP KVM.
No. A GPU has its own VRAM, separate from system RAM. Keep enough system RAM and NVMe speed to feed the GPU so it never waits on data.
No, VRAM is part of the GPU board. If you need more, move to a card with larger memory or a multi-GPU configuration.
Yes. GPUs render graphics and encode video for cloud gaming, headless game clients, VR and AR testing and live streaming. L4, A40 and L40S are the usual choices.
Start with the workload, then size VRAM, then check CPU cores, RAM and NVMe so nothing bottlenecks the GPU. Confirm your framework (CUDA, PyTorch, TensorFlow) supports the card, and plan for one more GPU if you expect to grow.
AI-powered · Instant results