Mellanox

Contamination is the leading cause of fiber link failure, and one dirty position on an MPO ferrule presents as a failing module. The inspect-clean-inspect discipline, the right tools, what never to do, and the symptom patterns that should send you to the scope instead of the spares shelf.

Before replacing a transceiver, read its DOM. Transmit and receive power split a dead link into three buckets in under a minute. Per-lane readings, the loss budget arithmetic, and the failures DOM is blind to.

One A40 48GB runs 27–34B models at 23–29 tokens/sec, fits a 70B at Q4 at 12–13.5 tokens/sec, and serves 1,705 tokens/sec of aggregate output on an 8B model at 100 concurrent requests. The measured numbers, the KV cache math behind the 72B cliff, and the three things an Ampere card cannot do.