One or two NVIDIA RTX 6000 with 24 GB VRAM per card, passed straight through to your instance.
The same base setup with different amounts of GPU power. The cards are passed directly through to your instance, so all the memory and the whole GPU are yours alone.
The starting point for image generation, rendering and models up to around 30B in INT4.
48 GB of combined VRAM has room for 14B in full FP16, 70B in INT4 and heavier render jobs.
Need more memory, more storage or a whole physical machine with a GPU? We will set it up.
The cards are not shared. Each instance gets the whole graphics card via PCIe passthrough, with all 24 GB VRAM and full compute available at all times.
The RTX 6000 combines 4,608 CUDA cores, tensor cores for AI and a dedicated video engine. That gives you one machine that handles graphics, video and models alike.
24 GB VRAM covers SDXL, Flux and ControlNet setups with plenty of headroom, without cutting resolution or batch size. Run your own ComfyUI or Automatic1111 on a machine only you have access to.
RT cores and 24 GB of memory keep large scenes with heavy textures on the card. Set up a render farm you can switch off once the project is delivered.
Dedicated NVENC and NVDEC engines turn transcoding into a GPU job instead of a CPU job. Ideal for streaming services, media archives and batch conversion.
Give your staff a powerful GPU machine wherever they are. The models stay safely in the datacenter, and the laptop only needs a browser or remote desktop (RDP).
Run 7B–14B models in FP16 or around 30B quantized to INT4 on a single card. Your own inference endpoint, without a single request leaving Norway.
The memory on the card decides which models you can run. Switch between one and two cards to see what fits.
The numbers are approximate and include weights plus normal KV cache usage. Actual needs vary with context length, batch size and the framework you use.
The Turing generation with 24 GB GDDR6 and ECC memory. A professional card built to run around the clock in a datacenter.
The card is handed to your instance in its entirety. No time slicing, no vGPU profiles and no other customers on the same memory.
ECC memory catches bit errors along the way. Important when a render job or training run has to go for many hours unattended.
576 tensor cores accelerate AI work, while 72 RT cores handle ray tracing in rendering. Full CUDA support for all common frameworks.
NVENC and NVDEC encode and decode video without loading the CPU, so transcoding can run alongside other work on the same machine.
Turing · TU102 · PCIe 3.0 x16
With two cards, both memory and compute double to 48 GB VRAM.
The GPU servers sit in our own datacenter in Sandefjord, and you always know where your images, models and customer data are.
A GPU server with us is first and foremost a proper server. Network, IP addresses and control panel are identical to our virtual and dedicated servers.
Symmetric port with 20 TB of included traffic, over our own network with multiple transit providers.
One IPv4 address and a full /48 IPv6 network come with every instance, at no extra cost.
Local NVMe SSD gives fast loading of models, textures and datasets straight from the machine.
Start, stop, console and invoices gathered in our own control panel, with an API for whatever you want to automate.
Take a snapshot before you experiment, and turn on daily backup.
Ubuntu, Debian, AlmaLinux, Windows or your own ISO. You install the drivers and frameworks you want to use.
Filtering in our own network keeps the service up if it should come under attack.
We support our customers directly from our offices in Sandefjord.
Can't find the answer you're looking for?
Send us a message and we'll get back to you quickly!
Get 10% discount on our Ryzen dedicated server Hobby with the code: HOBBYSOMMER