the 4th card needed a soldering iron
last measurement: 3 cards in pcie 1.1 x16 and the 4th one stuck in a 1.1 x4 slot. the x4 lane was holding the whole box back, so i pulled that card out and soldered it into a 1.1 x16 slot.
the difference
| prompt processing | generation | |
|---|---|---|
| 4× pcie 1.1 x16 | 270 t/s | 35 t/s |
before i soldered it in, that x4 card was the ceiling for the whole box. one card, one soldering iron, and the machine stopped being limited by a lane it outgrew a decade ago.
next
2 more cards are getting installed tomorrow. the plan, for now: tts and stt — or image generation, if they get bored. that makes 6 cards in one box, and at some point the closet stops being a closet and starts being a small datacenter.