vGPU servers split an NVIDIA H200 into dedicated VRAM slices. Compare the matching plans, model sizes and limits.
8 Dedicated Cores (AMD EPYC 9535) + NVIDIA H200 (7 GB VRAM) · 16 GB RAM · 512 GB NVMe · 2.5 Gbit/s
GB6: 2,210/12,400 · Steal: 0.0 %
12 Dedicated Cores (AMD EPYC 9535) + NVIDIA H200 (14 GB VRAM) · 32 GB RAM · 1 TB NVMe · 2.5 Gbit/s
GB6: 2,225/17,200 · Steal: 0.0 %
vGPU fits 8B–14B LLMs, embeddings, image generation and small fine-tunes. For 70B models you need a lot of RAM or multiple GPUs; a single 14 GB slice is not enough.
CPU servers are cheaper, but LLM inference without a GPU is slow. vGPU provides Tensor Cores and fast VRAM; VRAM is limited and the plans cost more.
Ollama runs on the vGPU server with the NVIDIA Container Toolkit. OpenWebUI provides the web interface. RS 2000 vGPU 7 is enough for 8B models; RS 4000 vGPU 14 for 14B.
The server matcher shows vGPU for AI and root servers for large models.
Check the voucher in the cart.
RS 2000 vGPU 7 with 7 GB VRAM is enough for Llama 3.1 8B in 4-bit quantization. For 14B models, RS 4000 vGPU 14 is better.
Not on a single 7/14 GB slice. 70B models need hybrid offload with a lot of RAM (e.g. RS 8000) or multiple GPUs; token rates are lower.
Check the code in the cart. The general exclusions (domains, VPS ARM, VPS Lite, merchandise) apply; vGPU plans may have their own terms.