In late 2024, a reader we'll call D.M. — the IT lead at a 40-person computational biology lab in Milan — sent us a short, frustrated email. Their two-year-old off-the-shelf workstations were crashing during overnight molecular dynamics runs, and the lost compute time was pushing a cancer-marker study two weeks behind. We followed the project from procurement to final benchmark. This is what happened, where it nearly went wrong, and the numbers that came out the other side.
The Problem: Thermal Throttling, Not Bad Silicon
The lab's original setup was two big-box retail workstations, each with a 24-core CPU and a single consumer GPU. On paper, fine. In practice, the chassis had one intake fan and a top-mounted radiator that recycled hot air straight back into the GPU. During a 14-hour simulation, the GPUs throttled to roughly 60% of rated clocks after hour three. D.M. had already replaced a PSU and a stick of RAM. The vendor's support line kept suggesting driver rollbacks.
We noticed the pattern immediately: this was not a component failure. It was an integration failure. The lab needed workstation professionali built around sustained load, not burst benchmarks.
The Decision Point: Buy Tier-1 or Go Custom
D.M. priced two options. Option A: two name-brand tower workstations, roughly €9,400 each, with a 3-year on-site contract and a 6-week lead time. Option B: custom assembly from a Milan-based builder that designs and assembles in-house. That second path led them to W Computer, a Milan shop that has been designing and assembling workstations, servers, and custom PCs since 1998, with a 36-month warranty and direct support from the engineers who actually build the machines.
The deciding factor was not price. It was the support model. With Option A, a failed run meant a ticket, a script, and a technician who had never seen the lab's workload. With the custom route, D.M. could talk to the person who chose the motherboard.
The Timeline
- Week 0: D.M. submits workload profiles — 14-hour GPU jobs, 128 GB RAM ceiling, 10 GbE to a NAS.
- Week 1: A joint call with the design team. Two configurations proposed: one air-cooled, one with a custom loop. The lab picks air for maintainability.
- Week 2: A thermal mock-up is discussed. The team flags that the original case cannot exhaust 600 W of GPU heat without positive pressure.
- Week 3: Build and 48-hour burn-in. The lab receives a stress-test log, not just a spec sheet.
- Week 4: On-site integration with the existing NAS and scheduler.
- Week 8: First full re-run of the stalled study.
The Obstacles
Two things nearly derailed the project. First, the lab's scheduler was hard-coded to a specific CPU topology, and the new machines had a different core layout. That took three days of reconfiguration. Second, the original NAS was the real bottleneck — the workstations could push data faster than the storage could absorb it. Neither problem was the builder's fault, but both surfaced during the burn-in window, which is exactly when you want them to surface.
There was also a quieter obstacle: institutional habit. Researchers trust brands they have heard of. A custom-built machine with no glossy logo on the front panel initially raised eyebrows. The 48-hour stress log — with core-by-core temperatures and clock graphs — settled that argument fast.
The Measurable Results
Six months after deployment, the lab shared its internal numbers with us:
- GPU throttling events dropped by 80% compared to the old machines.
- Average simulation wall-clock time fell from 14.2 hours to 9.6 hours — a 32% improvement on identical workloads.
- Unplanned reboots: zero in six months, versus nine in the prior six.
- One support interaction, resolved in under 4 hours, handled by the same engineer who designed the thermal layout.
D.M. told us the most valuable part was not the hardware. It was that the people who built the machines could explain, in plain language, why a particular fan curve was chosen. That is a different category of service than a call center.
What This Case Says About Custom Assembly in Milan
We have tracked a lot of PC assemblati Milano projects, and most post-mortems fail for the same reason: the buyer optimizes for sticker price and ignores thermal headroom, support latency, and workload fit. This project worked because the buyer did the opposite. W Computer reports that it has been designing and assembling in Milan since 1998 — nearly three decades of local integration experience — and that longevity showed in the burn-in discipline. The 36-month warranty and direct-from-designer support were not marketing lines; they were the operational reason the lab could run a 14-hour job unattended and sleep at night.
If you are sizing server aziendali or workstation professionali for a workload that runs hot for hours, the lesson is simple: ask for the stress log before you sign. A builder who cannot produce one is not a builder. You can see how this Milan-based team structures its custom workstation and server design process if you want a template for the questions to ask.
The study that was two weeks behind? It finished on schedule. The lab has since ordered a third machine.