We always appreciate feedback from our customers. One particular request kept coming up: a GPU server with multiple graphics cards. Now it's finally here: our first multi-GPU server GEX430.

The hardware inside the GEX430

Our new GEX430 multi-GPU server features four NVIDIA RTX PRO™ 6000 Blackwell Max-Q GPUs, each with 96 GB of GDDR7 ECC VRAM, providing a total of 384 GB. This makes the server particularly suitable for memory-intensive AI workloads, demanding rendering, simulations, and data science applications. It expands our GPU-Line for professional business applications.

The four GPUs occupy PCIe slots on the motherboard, each connected directly to the CPU via 16 PCIe 4.0 lanes. This gives each GPU access to the full bandwidth without having to share it with another GPU. With the right frameworks, you can either assign different tasks to individual GPUs or have all four work together on a compute-intensive workload.

Features:

Four NVIDIA RTX PRO™ 6000 GPUs with 96 GB GDDR7 ECC VRAM each, totaling 384 GB
4th-generation RT Cores
4 × 752 5th-generation Tensor Cores
AI performance: 4 × 3,511 AI TOPS
FP4 support
Intel® Xeon® Gold 5412U with 24 cores and 48 threads
2 × 1.92 TB Datacenter Edition NVMe SSDs (Gen4, software RAID 1)
128 GB DDR5 ECC registered RAM

Use cases

Running a large language model

Want to analyze extensive documents, tackle demanding coding tasks, or build complex AI agents? Then you might want to run a particularly large and powerful LLM.

Suitable models range from around 70 to 100 billion parameters at 16-bit precision (BF16). Their model weights alone require approximately 140–200 GB of VRAM. Using 8-bit precision (FP8) can roughly halve this requirement.

A good example is Kolibri by Aleph Alpha. Developed in Germany, this language model has 78 billion parameters and requires around 156 GB of VRAM at BF16 precision, so it won't fit on any single GPU in the server. With the right frameworks, you can distribute the model across multiple GPUs. The additional memory also leaves room for longer contexts – everything the model considers simultaneously when processing a request.

Another suitable example is Qwen3-Next-80B-A3B-Instruct, with 80 billion parameters and a theoretical memory requirement of around 160 GB at BF16 precision.

Serving many users

VRAM is also crucial for AI inference. The more requests a model processes simultaneously, the more memory it generally needs. With 384 GB, the GEX430 offers plenty of headroom to serve a large number of users concurrently.

This is useful for an online store's support chatbot when numerous customers suddenly reach out after a promotional campaign. It’s beneficial to implement an internal knowledge assistant when large numbers of your team want to use your model simultaneously during peak hours.

Running multiple LLMs simultaneously

Not every AI model excels at the same tasks. Some specialize in general-purpose applications, while others focus on programming, speech recognition, or document processing. For businesses, it can therefore make sense to run several models simultaneously. A large language model answers complex customer questions, a coding model supports the development team, and a smaller model sorts incoming requests.

With the GEX430, you can distribute these tasks across four different GPUs. This lets you run multiple models in parallel without having to reload them for every request, provided they fit into the available memory.

AI training, fine-tuning, and RAG

With the GEX430, you can adapt existing AI models to specific tasks. By fine-tuning your model, it can learn your company’s technical terminology, writing conventions, or specific workflows. Techniques such as LoRA reduce training effort by adjusting only a small portion of the model's parameters.

By combining the model with Retrieval-Augmented Generation (RAG), you can also enable it to answer questions based on your own knowledge base. It retrieves relevant documents and information while processing a request, without the need to retrain the model.

You can also train smaller models from scratch. Depending on the training method, you can distribute data or computations across multiple GPUs working in parallel. Practical use cases range from image recognition in quality control to models that automatically sort incoming emails.

3D rendering, video processing, and other use cases

The GEX40 also boosts 3D rendering. For example, compatible rendering software can use all four GPUs in parallel to render different frames of an animation. With a total of 384 GB of VRAM, there's also plenty of memory available for complex scenes with detailed textures.

The 4th-generation RT Cores accelerate calculations for lighting, shadows, and reflections directly in hardware. The server also delivers the power needed for video processing, scientific simulations, and data science.

Our GPU-Line keeps growing

This year, we gave our GPU-Line a complete refresh. All servers now feature modern GPUs, and we've doubled our lineup to four models. This gives you even more options to find the right server for your specific workloads.

You can also combine different servers. Not every model needs a total of 384 GB of VRAM. For example, smaller embedding models and speech recognition workloads can run on a GEX45 with 24 GB of VRAM, while the GEX430 handles more compute-intensive tasks. This lets you tailor your hardware to the specific requirements of your applications.

Your AI workload, your choice

Whether you need a multi-GPU or single-GPU solution, our GPU-Line offers you the server that's right for your AI workload.

Digital sovereignty

Many businesses want to use LLMs without sending sensitive data to third-party providers. With Hetzner, you can run a dedicated GPU server under your full control, on hardware located in the EU and hosted by a German provider. Your sensitive source code, customer data, and internal documents can remain within your own infrastructure if you properly configure your environment.

This makes it easier for you to meet data protection and compliance requirements. At the same time, you retain control over your models, data, and costs.