Measure Before You Procure

The bottleneck was system RAM, not VRAM. That single finding changed the procurement plan.

We assumed the GPU was the limiter. It wasn't. The bigger GPU, sitting on a server with 16 GB of system RAM, crashed at roughly 100 concurrent users. The smaller GPU, on a server with 64 GB of system RAM, held at full accuracy through about 500 concurrent users. Same model. Same workload. Different host memory.

We upgraded the production server to 128 GB and the problem went away. No new GPU. No vendor call. No quarter lost to a procurement cycle.

The lesson is not about hardware. It is about instrumenting the system before signing the purchase order.

Measure before you procure.