We finished moving the last Milan-generation hosts in late 2024. Enough time has passed to say something useful about what changed.
| Metric | Milan (7003) | Genoa (9004) | Δ |
|---|---|---|---|
| Geekbench 6 single | 1 674 | 2 310 | +38% |
| Geekbench 6 multi, 8 vCPU | 11 020 | 15 380 | +40% |
| 4K random read, QD32 | 186k IOPS | 312k IOPS | +68% |
| Memory bandwidth per socket | 204 GB/s | 460 GB/s | +125% |
| Idle power per host | 142 W | 189 W | +33% |
| Performance per watt, loaded | 1.00 | 1.34 | +34% |
The one that went the wrong way
Idle draw is a third higher. On a fleet that is never idle this is invisible; on the four sites where we keep standby capacity for failover, it added about €31,000 a year. We now consolidate standby capacity onto fewer hosts and accept a slightly longer failover, which is the right trade at that price.
What customers noticed
Almost nothing, which is the point. Live migration moved 94% of instances with no reboot; the remainder saw a pause under ninety seconds. Nobody paid more, and the price cut we made in 2024 came out of the density gain.