Back when I started in this business, a "server" meant 200W idle and nobody cared. Kids these days with their efficiency metrics. I care about one thing: TCO.
Replaced dual Xeon Gold 6354 (Ice Lake) with a single EPYC 9354 (Genoa) in March. Same workload—Proxmox cluster hosting ~80 VMs, mostly web and light DB. Power at the PDU:
- Ice Lake: 287W average
- Genoa: 172W average
That is 40% less at 0.14/kWh. Payback on the platform swap: ~14 months by my math. Faster than I expected.
But here is the thing. Mark my words: these efficiency gains are going to disappear when everyone crams AI inference onto the same silicon. Back when virtualization was the density play, we got sane power curves. Now? Wait for it.