Problem:

Power consumption

Current timing techniques are not efficient. Delay for different paths are not same. We design for the worst cases.

Razor: remove safe margin guarantee and recover from error.

Good chips: high frequency/low voltage (power) depends on market

High frequency v.s. low power: market need a guarantee frequency rather than “can run faster when timing margin is OK”.

Variations:

Manufactory Temperature Power supply Workload Aging

Users:

High-end workstation, server, HPC

Consider performance, power, and cost. Get as much as possible performance within certain constrains. This paper lock performance, try to reduce power.

Ways to reduce power:

Power regulator deal with long-term stuff (temperature, timing margin)

For short-term stuff, drop frequency to avoid soft failure. Use test circuits to determine delay.

Evaluation:

Less margin, less power. Minimum is 0 margin

For power, 0 margin and 50% margin are close.

There is no guarantee for “error free”, but the margin is tunable.