How we build

Consumer motherboards — the ones you find in a standard computer shop — are typically manufactured for 2 to 3 years. After that, the manufacturer discontinues them and no replacements are available. This is acceptable for a home computer. However, for a system running 24/7 in a factory, on a ship or in a server room, it poses a risk.

Industrial motherboards are manufactured with a minimum 7-year supply guarantee. The chipsets are selected for stability, not for the latest features. The temperature ranges are wider, the components are soldered (not socketed), and BIOS support continues for as long as the board is in production.

We use industrial motherboards from manufacturers such as Supermicro and ASRock Rack. For workstations where the latest chipset features are relevant (for example, PCIe 5.0 for NVMe), we use high-end consumer boards — but only if the application justifies it.

An Intel Z890 platform (Core Ultra 200 series) has 2 memory channels. That is sufficient for most workstation tasks, but it means that the total memory bandwidth is limited to around 90 GB/s.

AMD’s Threadripper platform (TRX50 chipset) has 4 memory channels — twice as many. For software that moves large amounts of data through memory — such as large After Effects compositions, Houdini simulations or scientific datasets — this makes a noticeable difference. The Threadripper Pro platform (WRX90) even features 8 memory channels, bringing its bandwidth close to that of the EPYC platform. In terms of PCIe lanes too, the Threadripper Pro platform can handle considerably more, with 128 lanes. So for computers that need to process a very large amount of data in the shortest possible time, Threadripper Pro is the recommended choice.

AMD EPYC (server-class) goes even further with 8 memory channels and support for up to 6 TB of RAM. This is the platform for serious computing tasks, databases and virtualisation.

The difference isn’t just bandwidth. More channels also mean more DIMM slots, so more scope for future expansion without replacing existing modules.

DDR5 memory with speeds exceeding the JEDEC specification (typically 4800–6400 MHz, depending on the platform) has been factory-overclocked. The manufacturer applies an XMP or EXPO profile to the module, which forces the CPU’s memory controller to operate outside its official specifications.

For overclockers and gamers who like to tinker, that’s fine. For a professional system that needs to run reliably day and night, it’s unacceptable.

Overclocking introduces variability. The memory may work 99.9% of the time — but that 0.1% could be a corrupted render, a crash halfway through a simulation, or quietly incorrect data in a calculation. We do not take that risk.

We always configure memory to JEDEC specifications. That is the speed guaranteed by Intel or AMD, which the memory controller can handle, and which will still function identically after 5 years.

ECC (Error-Correcting Code) memory adds an extra layer of protection: it automatically detects and corrects single-bit errors. For servers and workstations handling large datasets or running lengthy calculations, we always recommend ECC.

A GeForce RTX 5090 and an RTX Pro 6000 share the same GPU architecture. However, the drivers are fundamentally different. GeForce drivers are optimised for gaming: maximum frame rate, visual quality, and regular updates that can sometimes cause issues. Professional drivers (formerly Quadro) are optimised for stability and are certified by software vendors (ISV certification).

What does that mean in practice?

Use GeForce (RTX 5080/5090) if:

• Your software uses GPU rendering (Redshift, Octane, Blender Cycles, DaVinci Resolve)
• You need CUDA computing power (AI training, scientific simulation)
• Maximum VRAM per euro is the most important factor

Use RTX Pro if:

• You work with CAD software that requires ISV certification (SolidWorks, CATIA, Siemens NX)
• You require 10-bit colour rendering via the GPU (rather than via a Blackmagic card)
• Your organisation requires certified hardware for support contracts

GeForce cards generally offer more processing power and VRAM per euro. The Pro cards offer stability and certification. We advise based on your software and workflow, not on profit margin.

For AI applications, the Pro cards are almost always better. Only in image generation does the chip become important, whereas VRAM is essential. And it is precisely in the area of VRAM that the RTX Pro 6000 card, with 96GB of VRAM, is unbeatable.

A modern Intel Z890 platform has 24 PCIe lanes available (20 from the CPU + 4 from the chipset). That sounds like a lot, but let’s do the maths:

• GPU: 16 lanes (x16 slot)
• First NVMe SSD: 4 lanes
• Second NVMe SSD: 4 lanes

That’s 24 lanes — all used up. There’s no room left for a 10GbE network card, an HBA for extra storage, a capture card or a second GPU. Every addition means something else has to be sacrificed, or that the GPU drops back to x8 (half the bandwidth).

AMD Threadripper (TRX50) offers 48 PCIe lanes and the WRX90 chipset can even handle 80 PCIe lanes. AMD EPYC offers up to 128 lanes. On those platforms, you can run a GPU at x16, four NVMe drives, an HBA, a 10GbE card and a GPU for compute side by side without compromise.

This is why a ‘fast’ consumer CPU does not automatically make for a good workstation. The CPU may be fast — but if the rest of the system is bottlenecked, it doesn’t matter.

NVMe Gen4 vs Gen5

In benchmarks, NVMe Gen5 SSDs achieve 14 GB/s compared to 7 GB/s for Gen4. In practice, you will only notice this difference during sustained sequential reads of very large files. For most workloads — booting up, loading applications, opening project files — Gen4 is more than fast enough and considerably cheaper.

We only recommend Gen5 if your workflow has been proven to benefit from the extra bandwidth, for example when editing uncompressed 8K video.

HBA cards

An HBA (Host Bus Adapter) is a card that provides direct access to storage media without the intervention of a RAID controller. In combination with ZFS or a similar filesystem, you have full control over redundancy, checksumming and data integrity at the software level.

Why not use hardware RAID?

Hardware RAID controllers add complexity without providing any added value. They have their own firmware that can fail, their own battery-backed cache that can run flat, and they lock your drives to a specific controller. If the controller breaks down, you will need another unit of exactly the same model to read your data.

Software RAID (ZFS, mdadm) is transparent, portable and controllable. We always recommend software-based solutions for data integrity.

IPMI / BMC (out-of-band management)

Every serious server has a BMC (Baseboard Management Controller) — a small independent computer on the motherboard that is always on, even when the server itself is switched off. This allows you to switch on the server remotely, configure the BIOS, install the operating system and diagnose problems. Without IPMI, you are dependent on physical access — and that is not an option for a server in a data centre or server room.

ECC memory

Servers run 24/7, often for months without a reboot. During that time, bit flips inevitably occur due to cosmic radiation and electrical noise. ECC memory automatically detects and corrects these errors. Without ECC, a server quietly accumulates errors in its memory — which can lead to corrupted data, crashes or security issues.

Redundant power supply

A server with a single power supply has a single point of failure. With dual (redundant) power supplies, the system continues to run if one fails. Combined with a UPS and monitoring, this forms the basis for uninterrupted availability.

What we build

Our servers are based on Supermicro platforms with IPMI, ECC memory and optional redundant power supplies. We build file servers, backup servers, virtualisation hosts and dedicated compute nodes. Each system is configured and tested for the specific task for which it is intended.