Components

A GPU build is planned backwards

Starting from the case and the CPU feels natural, but for a compute machine the order is different. The cards come first, because they set everything else: how many PCIe lanes the processor needs, how much system memory to fit, which power supply survives the peaks and whether it all fits the case.

Then comes the bottleneck check. The fastest card will sit idle if the CPU cannot prepare data fast enough, the drive reads the dataset in bursts, and there is less memory than the model needs to unfold.

The order this is usually worked out in

• Graphics cards. How many, how much memory each, what cooling. Everything downstream follows from this.
• CPU and platform. The numbers that matter are PCIe lanes and memory channels, not clock speed.
• System memory. A rule of thumb for GPU machines: about twice the total GPU memory.
• Storage. A mirrored M.2 pair for the system, large U.2 drives with high write endurance for data.
• Power and case. Wattage headroom for card spikes and airflow that actually carries the heat out.

How we can help

We build these machines daily and see where configurations break. The usual cases: a CPU short on lanes, a power supply sized to the edge, two memory sticks on an eight-channel platform.

Tell us what you plan to run and within what budget. We will put together a list where every part is there because it is needed, not because it costs more.