The example fleet
Take 10,000 cards averaging 75 W each, running all year, in a facility with a power usage effectiveness (PUE) of 1.2. The energy a card draws is paid for again in cooling and distribution, which is what PUE captures.
10,000 cards × 75 W = 750 kW of card power. Over 8,760 hours that is 6,570 MWh a year on the cards themselves, and 7,884 MWh at the wall once PUE is applied.
| Reduction in card power | Fleet energy saved per year |
|---|---|
| 1% | 79 MWh |
| 5% | 394 MWh |
| 10% | 788 MWh |
Where the reduction comes from
Vendor timing models cover the worst chip at the hottest corner at the end of its life, so every card runs at a voltage chosen for that chip. Measured per chip, the margin needed shrinks to measurement error plus drift until the next measurement; each link is resolved to under 20 ps. The voltage that margin no longer needs is the reduction.
How large it is depends on the margin your chips carry today, which is why an evaluation starts by measuring a sample of your own cards before anything is recommended. On production shells at realistic utilization we measure 30–54% of timing margin carried for the worst-case corner.*
Substitute your own numbers
- Card count and average power from your inventory and your power telemetry.
- Hours: 8,760 for cards that never sleep; fewer if they do.
- PUE from your facility. Anything above 1.0 multiplies what you save on the card.
- The reduction from an evaluation on your cards, not from this page.
Safety is part of the arithmetic
A change that risks timing is not a saving. Every change is scheduled, handshake-gated and measured again within a second, and the aging model pulls the operating point back before a chip drifts toward the threshold. Power and performance explains the guardrails.
*Fluid Silicon measurements. They vary with variation pattern, workload, vendor and device age.