Abstract
The present disclosure relates to systems and methods for managing power consumption and reducing power fluctuations in data centers utilizing compute accelerators. A two-threshold power capping scheme applies differing levels of throttling to computational tasks based on aggregated power usage. An adaptive throttling multiplier dynamically adjusts the upper bound of power consumption based on historical power spike events. Furthermore, an adaptive hardware power slope control mechanism adjusts guaranteed power floors based on anticipated computational loads, extending the operational lifespan of the hardware slope control circuits while preserving application performance. Asynchronous processing threads execute power state reconciliations to minimize latency and ensure operational efficiency. Keywords: power capping, load shedding, compute accelerators, dynamic throttling, power smoothing, data center management, asynchronous processing.
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Recommended Citation
N/A, "Smart Throttling and Power Management for Compute Accelerators", Technical Disclosure Commons, ()
https://www.tdcommons.org/dpubs_series/9612