# GPU server setup help from our engineers > GPU server setup help included with every server: first-boot checks, environment and migration advice from our engineers, and sizing tools before you order. Source: https://overgpu.com/products/expert-onboarding Updated: 2026-10-22 Publisher: OverGPU (https://overgpu.com/) ## Key points - **Sizing before you order:** The VRAM calculator, GPU pages and docs show the memory math for your model, precision and workload. - **First-boot checks:** Confirm every GPU with nvidia-smi and, on 8-GPU servers, test the GPU interconnect with NCCL. - **Environment advice:** Drivers, CUDA or ROCm, containers and frameworks: advice on getting your stack running on your image. - **Data migration advice:** How to move datasets and checkpoints to your new server. Inbound traffic is not counted against your bandwidth. - **24/7, no extra fee:** Part of the support included with every server, through tickets in your dashboard once it is delivered. - **Your server, your control:** This is help and advice. You keep full root, and we do not operate your workloads. ## Recommended configurations - [NVIDIA H200](https://overgpu.com/gpus/nvidia-h200): $3,090 per month — 141 GB: a 70B model in FP8 with room for the KV cache - [8× NVIDIA HGX H100](https://overgpu.com/gpus/nvidia-h100): $16,900 per month — Eight H100s on NVLink, with help running NCCL tests at handover - [NVIDIA B200](https://overgpu.com/gpus/nvidia-b200): $4,590 per month — 180 GB: a 70B model in BF16 on a single GPU ## Overview GPU server setup help is part of every OverGPU server. As part of our 24/7 support, our engineers help you check the server after handover and get your environment and data in place, at no extra fee; before you order, our calculators and guides help you size it. It is help and advice: you keep full root and full control, and we do not run your workloads for you. ## What GPU server setup help includes Onboarding covers three moments: before you order, when your server is delivered, and while you move your work in. Before you order, the help center, the docs, the GPU pages and the calculators answer sizing questions; once your server is delivered, an engineer helps through support tickets in your dashboard, at any hour. | When | What we help with | | --- | --- | | Before you order | Choosing the GPU and the GPU count from your model and workload, with the VRAM calculator and our guides | | At handover | First-boot checks: GPUs, topology, containers, and NCCL tests on 8-GPU servers | | Moving in | Drivers, CUDA or ROCm, containers, frameworks and data migration | ## Sizing before you order Most of the cost is decided when you choose the GPU. Before you order, work out: - the model, its parameter count and the precision you plan to use; - whether you train, fine-tune or serve, and for serving, the context length and concurrency you expect; - the size of your dataset and how long you need the capacity. The [VRAM calculator](https://overgpu.com/tools/vram-calculator) turns these into a memory estimate, and each GPU page lists what fits. A typical result: a 70B model needs about 70 GB in FP8, so one H100 is tight and one H200 is comfortable. In BF16 it needs about 140 GB, which makes an H200 tight and a B200 or B300 comfortable. Those figures cover the weights; add the KV cache and plan 10 to 30% headroom. Full fine-tuning with Adam needs roughly 16 to 20 bytes per parameter before activations, which turns a 70B model into a multi-GPU job. When you [deploy](https://overgpu.com/deploy), describe your workload in the notes for our engineers. ## First-boot checks after handover When your server is delivered and its IP address and SSH details appear in your dashboard, run a few checks before real work starts. We help you read the output. ``` nvidia-smi nvidia-smi topo -m docker run --rm --gpus all ubuntu nvidia-smi ``` The first confirms that every GPU is visible with its full memory. The second shows the interconnect, which on HGX servers means NVLink between all eight GPUs. The third confirms that containers can reach the GPUs. On 8-GPU servers, an NCCL all-reduce test puts the interconnect under load. Build NVIDIA's nccl-tests and run, for example: ``` ./build/all_reduce_perf -b 8 -e 8G -f 2 -g 8 ``` If a GPU is missing or the bus bandwidth looks low, send us the output in a support ticket. Hardware faults are ours to fix: failed GPUs and components are replaced as a priority. On AMD MI355X servers, the equivalents are `rocm-smi` and RCCL tests. The [multi-GPU and NCCL guide](https://overgpu.com/docs/multi-gpu-nccl) has more on these tests. ## Environment and data migration advice Open a support ticket about the layer between the image and your code: - **Drivers and CUDA.** Which driver branch your framework expects, and whether a container's CUDA version suits the host driver. - **Containers.** GPU containers with Docker and the NVIDIA Container Toolkit, or ROCm containers on AMD. - **Frameworks.** Getting PyTorch, vLLM, SGLang or TensorRT-LLM to use every GPU, and the settings that matter for multi-GPU runs. For data migration, we advise on moving datasets, checkpoints and model weights from another provider or from an earlier OverGPU server: `rsync` over SSH for directories, parallel streams for large transfers, and checksums before you delete the source. Inbound traffic is not counted, so uploading to your new server does not use your included bandwidth. If your current provider bills egress, plan for that cost on its side. Our [PyTorch and vLLM guide](https://overgpu.com/docs/pytorch-vllm) and the [OS images](https://overgpu.com/products/os-images) page cover the software side. ## What GPU onboarding is not Onboarding is help and advice, not managed operations. We do not: - run, monitor or restart your training jobs or inference services; - maintain your application code, models or containers; - back up your data, unless agreed in writing; - access your data, except to operate or repair the service at your request, or when the law requires it. What we do own is the infrastructure: power, network and hardware, under a 99.9% monthly availability SLA with service credits, and the 24/7 support behind it. If a GPU fails, that is ours to fix. Your workloads stay yours to run. ## How to get setup help Before you order, answers are in the [help center](https://overgpu.com/help), the docs, the GPU pages and the calculators, and when you order, the notes for our engineers can describe your workload; once your server is delivered, open a support ticket from your dashboard, and an engineer answers at any hour. Include your server, what you are trying to do and any command output. There is no separate fee and nothing to book. Our page on [getting support](https://overgpu.com/help/getting-support) explains what to include in a request. ## Frequently asked questions ### Does setup help cost extra? No. It is part of the 24/7 support by engineers included with every server. ### Will you set up and run my training jobs or inference service? No. We give help and advice on sizing, drivers, containers, frameworks and data migration, but you run your workloads with full root access and we do not operate them for you. ### What first-boot checks should I run? Run nvidia-smi to confirm every GPU and its memory, and on 8-GPU servers run NCCL tests to check the interconnect. Send us the output in a support ticket if anything looks wrong. ### Can you help me move data from another provider? Yes, with advice on tools and approach, such as rsync over SSH and checksums. Inbound traffic is not counted against your bandwidth allowance. ### How do I reach an engineer? Through a support ticket in your dashboard once your server is delivered; before you order, answers are in the help center, the docs, the GPU pages and the calculators. Engineers answer tickets 24/7, so include your server, your goal and any command output. ## Related pages - [GPU server quickstart](https://overgpu.com/docs/quickstart): Log in over SSH, check the GPUs, update Ubuntu and run a first GPU container, with a check at each step. - [NCCL tests on a multi-GPU server: all_reduce_perf step by step](https://overgpu.com/docs/multi-gpu-nccl): Check NVLink topology, build nccl-tests, read busbw from all_reduce_perf and launch torchrun jobs across 8 GPUs. - [LLM VRAM calculator: will your model fit on a GPU?](https://overgpu.com/tools/vram-calculator): Pick Llama, Qwen, Mistral or gpt-oss, set the precision and context length, and see weights, KV cache and which GPUs fit. - [GPU server support: how tickets work](https://overgpu.com/help/getting-support): GPU server support at OverGPU runs 24/7 through dashboard tickets handled by engineers. See who can open one, what to include and what needs no ticket. - [GPU server OS images with CUDA, ROCm and Docker](https://overgpu.com/products/os-images): Ubuntu 22.04 and 24.04 with NVIDIA drivers, CUDA, Docker and the Container Toolkit, plus ROCm for AMD. See what is preinstalled. - [How to choose a GPU server for AI](https://overgpu.com/blog/how-to-choose-a-gpu-server): A framework for memory, bandwidth, precision, interconnect, CPU and storage, software and billing, plus a checklist.