Supermicro X13 4U Universal GPU SuperServer

A flexible 4U multi-GPU platform for AI inference, training and HPC, built so the accelerator choice is yours rather than locked to one vendor. The on-premise compute tier for serious AI workloads.

Price on request · Talk to a Sirap engineer or call (+356) 2138 5911.

Sirap, your Malta partner for Supermicro, with local specification, installation, service and support.

What the Universal GPU SuperServer is for

This is the workhorse GPU server for on-premise AI: a 4U platform that takes multiple data-centre GPUs and the dual Intel Xeon processors and fast storage to keep them fed. “Universal” is the important word, the design supports a flexible mix of accelerators and interconnects rather than tying you to a single GPU vendor, so the machine can be specified to the workload and re-specified as your models change. For a Maltese operator running inference in production, or training models in-house, this is the core compute tier described in our guide to on-premise AI compute.

Who it is for

Regulated finance, fintech and iGaming operators running fraud, AML, risk or personalisation models on-premise, and research or analytics teams needing serious acceleration on hardware they control. Where the requirement is the very heaviest NVIDIA-based AI work, the X14 5U NVIDIA GPU server is the step up; where a single GPU at a branch is enough, the AS-3015TR-i4 tower is the proportionate choice. A GPU server is only as good as the storage feeding it, which is where DataCore and the storage server come in.

Key features

Vendor-flexible acceleration. The Universal GPU design supports a mix of GPU and interconnect options, so the accelerator is matched to the workload and not dictated by the chassis.

Multiple GPUs, properly fed. Dual Intel Xeon processors, DDR5 and PCIe 5.0 provide the host bandwidth and memory to keep several accelerators supplied rather than waiting.

Built for sustained load. Server-grade power, cooling and airflow designed for GPUs running at full utilisation for long periods, which is what production inference and training actually do.

Storage-aware design. Fast local NVMe and high-speed networking so the data path to the GPUs is not the bottleneck, the most common reason an on-prem AI build disappoints.

Technical specifications

GPU type and count, memory and networking are specified to the workload, so this is quoted as a configured build, not a fixed SKU. Sirap confirms the exact configuration against the current Supermicro datasheet and your AI workload at quote.

What is in scope from Sirap

A GPU server is the part of an AI build most often over- or under-specified, and getting it right is the work we do: matching the GPU count and type to the real inference or training load, designing the storage and network path so the accelerators stay fed, planning the power and cooling for a load that is far heavier and peakier than ordinary servers, and supporting it from Malta. For regulated operators we configure it to sit inside your continuity and access controls from day one, not as an afterthought.

Often paired with

• DataCore storage on X13 rackmount nodes, plus an X13 storage server for the data lake behind the GPUs.
• Three-phase Eaton power sized for real GPU draw, with the certified Network-M3 card for NIS2 and DORA, covered in our UPS cybersecurity guide.
• The X14 5U NVIDIA GPU server when the workload needs the heaviest NVIDIA platform.

Could this be grant-funded?

On-premise AI infrastructure can align well with Malta Enterprise schemes aimed at digital and productive investment, such as the Smart & Sustainable Investment Grant or Digitalise Your SME. Fit depends on the project. Read our guide and see if your project qualifies, or talk to a Sirap engineer.

Frequently asked questions

What does “Universal GPU” actually mean?
That the platform is designed to take a flexible mix of GPU and interconnect options rather than being built around one vendor’s accelerator. It lets us specify the right GPU for your workload and re-specify later as your models change, without replacing the server.

How many GPUs do we need?
Fewer than most people assume for inference, more for training. We size the accelerator count to your real workload with sensible headroom rather than to a benchmark, because an over-populated server that sits half-idle is wasted capital and power.

What stops the GPUs sitting idle waiting on data?
The storage and network design. A GPU server needs fast local NVMe and a high-speed path to shared storage, which is why we design the DataCore or storage tier and the fabric alongside the server, not after it.

Talk to a Sirap engineer

Tell us the AI workload, whether it is inference, training or both, and your data and power constraints, and we will specify the GPU server, design the storage and power around it, and support it from Malta.

Request a quote · Price on request