Home lab serverPart 10 of 13

Build note

Why the Tesla P40 stayed out of this build

The Tesla P40 has 24 GiB of VRAM, the same capacity as the RTX 3090. That did not make it a useful third card for this machine. Memory capacity is only one part of what an inference card needs.

The P40 is a Pascal GPU. It has no tensor cores, and its FP16 path is a poor fit for the fast half-precision workloads I wanted to run. The card’s memory looked attractive, but the compute features and software support were the limiting factors.

The software stack was the first filter

The inference stacks I was considering, including vLLM and SGLang, expected compute capability 7.0 or newer. The P40 is compute capability 6.1. That meant the card was outside the supported path for the workloads I wanted to serve.

Even where a backend can run on Pascal, the result is not automatically useful. For this P40, FP16 throughput was about one sixty-fourth of its FP32 rate. A 24 GiB memory pool does not help enough if the math path is too slow or the serving software will not use the card.

One machine cannot use incompatible driver branches

There was also a driver conflict. Pascal support ended on an older driver branch, while the other workstation plan required a newer branch for the RTX cards and the GPU-unlock setup. The kernel loads one NVIDIA driver stack; mixing incompatible branches in one host was not a practical way to keep both sets of cards working.

The P40 also uses passive cooling. It needs airflow across the heatsink, so installing it in a workstation without checking the cooling path would add a thermal problem on top of the software one.

The decision

I left the P40 out of this build. If I use it later, it belongs in a separate airflow-served box with a software stack selected for Pascal, rather than as a third card beside the RTX 3090 and RTX 3060.

The rule that would have prevented the next week

Do not buy or install a card based on VRAM alone. Check compute capability, precision support, the serving framework’s minimum architecture, driver compatibility and cooling before planning the model around it.

How this was made

  • Written by: Rupesh, from his GPU compatibility investigation and P40 decision, with AI drafting and editing.
  • Compatibility details reflect the software versions and driver branches checked during that investigation; they are not a statement about every future release.
  • AI-assisted sections: structure and copy editing.