Everything we check before we quote

The processor and board are different sockets. The chip will not seat.

Quantity exceeds what the board physically has.

Multi-socket systems need identical stepping, core count and TDP. Mismatched CPUs halt at POST.

The P-series parts are locked at the factory to refuse working in pairs, even though they look identical.

Half the DIMM slots and several PCIe slots are wired to CPU2 and will not function.

The chassis has standoffs for other form factors.

No rails, no ears, no enclosed airflow path. It cannot go in a cabinet.

40mm fans at high static pressure are 60 dBA and up under load, exhausting into whoever sits behind.

Every failed boot becomes a site visit.

None of the selected processors ships with a heatsink.

Every populated socket needs its own.

Server sockets in particular have almost no aftermarket bracket options.

Rated below the processor's sustained peak. Escalates from warning to blocking past 25% short.

The panel or lid will not close.

Larger than the chassis accepts.

Rack heatsinks have no fan and do nothing in a tower.

Different keying, it cannot be inserted.

Registered memory will not train on a consumer controller, and unbuffered will not initialise on a server one.

The controller falls back to a slower interleave. On a 12-channel EPYC, eight modules costs roughly a third of memory bandwidth.

They run, but correction stays off and single-bit errors pass silently.

Intel restricts ECC to Xeon and W-series.

Anything over about 40mm fouls the cooler's front fan.

You are paying for speed the platform cannot use.

These mount to a vendor baseboard sold as a complete system. There is no PCIe edge connector.

Datacenter accelerators have no fans and throttle within a minute in a tower.

Triple-slot coolers consume neighbouring positions even though they use one PCIe connector.

It works, but host bandwidth is halved. Visible in multi-GPU training, not in gaming.

GPUs, network and NVMe together exceed what the processor provides.

Data-parallel training runs at the pace of the slowest card and is bounded by its VRAM.

Bridges pair cards two at a time; one ends up unbridged.

A second identical card and a bridge gives one pooled memory space.

U.3 backplanes accept U.2, SATA and SAS. U.2 backplanes accept U.2 only, the connector mates but a U.3 drive will not enumerate.

Without power-loss protection an outage can lose already-acknowledged writes.

30mm longer than 2280. Many desktop boards have no standoff for it.

Right for the read-heavy dataset tier, wrong for checkpoints and scratch.

The unit shuts down under load, usually mid-job.

Running a 450W-plus card off a four-way 8-pin adapter is the most common cause of melted connectors.

Daisy-chaining two GPU connectors onto one cable exceeds its 150W rating.

An empty header causes shutdowns under all-core load, it passes a quick test and fails a real job.

Pakistani single-phase mains is 230V/50Hz.

Running near the ceiling shortens life, maxes the fan and leaves nothing for expansion.

Pre-3.0 units trip their over-current protection on the transients these cards produce.

A 16A Pakistani domestic circuit is 3.68kW; continuous load should stay under about 2.9kW.

Keyed differently from PCIe 8-pin. Plugging a PCIe cable in damages the card.

Suits Pakistani mains; will not start on 110V.

The cable is identical and the connectors mate, but the protocols are unrelated.

Cage sizes are not interchangeable.

Most switches read the module EEPROM and refuse to bring the port up. Electrically fine, firmware says no.

Ports ship empty; optics are a meaningful share of the fabric budget.

You keep RoCE and GPUDirect, you lose in-network reduction.

Every gradient exchange bounces through host memory.

You need roughly 150mm behind the equipment for cable bend radius and the vertical PDU.

An overloaded UPS drops to bypass, doing nothing at the moment it is needed.

The doors become the thermal restriction.

Beyond about 7.4kW a single-phase 32A feed is exhausted.

Given local grid behaviour, double-conversion is the safer specification.

Load-shedding will otherwise cut a running job and risk the filesystem.